An enterprise digital asset management method based on a knowledge graph

Abstract

The application discloses a kind of enterprise digital asset management methods based on knowledge graph, specifically includes: the initial graph is built with asset entity as node, multivariate dependence as hyperedge;Each hyperedge time series causality test is deleted, and the sparse hypergraph is obtained by deleting the hyperedge below threshold value;Each hyperedge is split into multiple relationship subgraphs by multi-head attention;Parallel input space aggregation unit of space-time decoupling network, output space aggregation embedding;Again, time aggregation unit outputs time aggregation embedding;Each asset node health score and risk early warning are obtained through classification layer;The node and associated hyperedge below threshold value are marked as abnormal cluster, and output topology structure graph.The application improves risk early warning accuracy by hyperedge pruning and multi-head decoupling.

Description

Technical Field

[0001] This invention relates to the field of knowledge graph asset management technology, and in particular to a knowledge graph-based method for enterprise digital asset management. Background Technology

[0002] Enterprise digital asset management involves the unified inventory, relationship tracking, and health assessment of digital assets such as software, data, APIs, and configuration items. Existing asset management methods mainly rely on configuration management databases (CMDBs) and metadata management platforms, collecting asset information through manual input or scripts, and storing the dependencies between assets in relational tables or key-value pairs.

[0003] This type of method has two technical drawbacks. First, it only records direct connections between assets, failing to automatically discover hidden impact paths across systems and involving multiple hops. When underlying assets change, managers struggle to predict which upstream businesses will be affected, leading to reactive fault response. Second, the inconsistent data formats of heterogeneous assets from multiple sources—documents, code, database tables, API interfaces, etc.—stored scattered across different systems, lacking a unified semantic framework, makes automated asset valuation and risk transmission analysis difficult.

[0004] Some existing technologies attempt to introduce knowledge graphs to organize asset data, but their relationship modeling is still limited to paired edges, failing to express multi-dependencies involving more than three assets (such as a compliance audit simultaneously constraining multiple data tables). Furthermore, existing graphs are mostly statically constructed, lacking the ability to model asset temporal changes and causal propagation. Therefore, there is an urgent need for an asset management method that can support multi-relationship expression, dynamic pruning, and causal reasoning. Summary of the Invention

[0005] One objective of this invention is to propose a knowledge graph-based enterprise digital asset management method. This invention introduces a superedge pruning mechanism and a temporal causal decoupling model to perform dynamic sparsity filtering and relation type splitting on multi-dependency superedges in the asset knowledge graph, and constructs a continuously optimized spatiotemporal aggregation evaluation process. It has the advantages of high accuracy in discovering latent influence paths, fast efficiency in cross-system dependency tracking, and stable asset health early warning response.

[0006] A knowledge graph-based enterprise digital asset management method according to an embodiment of the present invention includes the following steps: Acquire multi-source heterogeneous data of enterprise digital assets, construct an initial asset knowledge graph, and use asset entities as nodes and multi-dimensional dependencies between assets as hyperedges; The causal effect strength of each hyperedge to key asset nodes is calculated by the temporal causal test method. Hyperedges with causal effect strength below a preset threshold are deleted to obtain a sparse hypergraph. Each hyperedge in the sparse hypergraph is input into the multi-head attention decomposition module. Based on the attention weights of the nodes within the hyperedge for different relation types, each hyperedge is split into several relation type subgraphs, and each relation type subgraph corresponds to an asset relation semantic. Several relation type subgraphs are input in parallel into the spatial aggregation unit of the spatiotemporally decoupled heterogeneous graph network. The spatial aggregation unit aggregates the features of neighbor nodes along the hyperedge structure and outputs the spatial aggregation embedding. Spatial aggregation embedding is input into the temporal aggregation unit of the spatiotemporally decoupled heterogeneous graph network. The temporal aggregation unit aggregates the embedding sequence of several historical moments along the time axis and outputs the temporal aggregation embedding. By embedding time aggregation into the fully connected classification layer, a health score and risk warning signal for each asset node can be obtained. Mark the asset nodes whose health scores are below the warning threshold and their associated superedges as abnormal asset clusters, and output the topology diagram of the abnormal asset clusters.

[0007] Optionally, the multi-source heterogeneous data of the enterprise's digital assets specifically includes: Asset configuration records, asset metadata definitions, descriptions of dependencies between assets, asset change history logs, asset usage statistics, technical documentation text associated with assets, source code structure information corresponding to assets, and interface definition data provided by assets to the outside world.

[0008] Optionally, obtaining the sparsed hypergraph includes the following steps: Collect the time-series feature vector of each asset node within a continuous time window. The time-series feature vector includes asset change frequency, asset call response time, asset error rate, and the rate of change of the number of upstream and downstream connections of the asset. Select a set of critical asset nodes, which includes asset nodes that carry core business processes, convergence nodes that aggregate more than five dependent paths, and asset nodes that have appeared three or more times in historical fault records. For each hyperedge, extract the temporal feature vector sequence of all nodes within the hyperedge within the time window, input the sequence into the structure vector autoregressive model, and estimate the conditional Granger causality test statistic of each node within the hyperedge to the key asset node. Set a causal strength threshold and define the causal strength of each hyperedge as the weighted average of the Granger causality test statistics of all nodes within the hyperedge to the key asset node. The weights are determined by the inverse of the graph distance between the nodes within the hyperedge and the key asset node. Remove hyperedges with causal strength below the causal strength threshold to obtain a preliminary sparsified hypergraph; The temporal feature vectors of the remaining hyperedges in the initial sparsified hypergraph are collected, and the historical causal evolution trajectory of each remaining hyperedge is constructed. The causal evolution trajectory contains the causal intensity sequence of three time windows. The causal evolution trajectory is input into the long short-term memory network to predict the causal intensity decay coefficient of each remaining superedge in the next time window; Set an upper limit for the decay coefficient. When the predicted decay coefficient is lower than the upper limit, retain the hyperedge. When the predicted decay coefficient is not lower than the upper limit, delete the hyperedge to obtain the final sparse hypergraph.

[0009] Optionally, splitting each hyperedge into multiple relation-type subgraphs includes the following steps: For each hyperedge, construct a node feature matrix and an edge type encoding vector. The rows of the node feature matrix correspond to the embedding vector of each asset node within the hyperedge, and the columns correspond to the embedding dimension. The length of the edge type encoding vector is equal to the total number of predefined relation types. Each position takes a value of 0 or 1, where 1 indicates that this hyperedge contains a priori indication of the corresponding relation type. The node feature matrix and edge type encoding vector are input into the multi-head attention decomposition module. The multi-head attention decomposition module contains multiple parallel attention heads. Each attention head independently calculates a set of attention weight matrices. The elements of the attention weight matrix represent the association strength between nodes within the hyperedge under a specific relation type. Among the multiple parallel attention heads, some attention heads calculate the structural similarity weight by multiplying the transpose of the node feature matrix, another part of the attention heads calculates the semantic compatibility weight by multiplying the node feature matrix and the learnable type embedding matrix, and the remaining attention heads calculate the evolutionary consistency weight by multiplying the temporal difference features of the node feature matrix and the topological centrality weight by multiplying the degree vector of the node feature matrix, respectively. After summing the element-wise output weight matrices of all attention heads, the result is input into the softmax function to obtain the normalized relation assignment probability matrix. Each column of the relation assignment probability matrix corresponds to a predefined relation type, and each row corresponds to a node inside the hyperedge. For each predefined relation type, extract the node pairs whose probability values ​​in the column of that type in the relation assignment probability matrix exceed the assignment threshold, use these node pairs and their corresponding probability values ​​as edge weights, and construct the subgraph corresponding to that relation type. The subgraph only contains nodes whose probability values ​​in the nodes inside the hyperedges exceed the assignment threshold. Delete subgraphs with fewer than 2 nodes, and retain subgraphs with 2 or more nodes as valid relation type subgraphs.

[0010] Optionally, the asset relationship semantics corresponding to each relationship type subgraph are determined by a predefined set of relationship types, which includes the following eight asset relationship semantics: Physical dependency semantics indicate that the operation of one asset instance requires computing or storage resources provided by another asset instance; Logical derivation semantics indicates that the data content or functional definition of one asset originates from another asset; The semantics of time-triggered relationships indicate that a change event of one asset will trigger another asset to perform a preset operation within a specified time window; Governance constraint semantics indicate that one asset must comply with the security policies or compliance rules defined by another asset. The semantics of value transmission relationships indicate that changes in the effectiveness of one asset will propagate along the graph and affect the business contribution of another asset; Redundancy and backup relationship semantics indicate that two assets are functionally interchangeable, and the other takes over in the event of a single point of failure; Version inheritance semantics indicate that one asset is a historical version or iterative derivative version of another asset; Access authorization relationship semantics, indicating that one asset holds the authorization credentials to read or write to another asset; The multi-head attention decomposition module outputs eight sets of relation type subgraphs for each hyperedge. Each set of subgraphs corresponds to one of the eight asset relation semantics mentioned above. The node pairs retained in the subgraphs represent asset node pairs within this hyperedge whose association probability value under this relation semantic exceeds a preset allocation threshold. When a superedge is decomposed by multi-head attention, if the number of nodes in three or more of the eight subgraphs is zero, this superedge is marked as a noisy superedge and deleted, and will not participate in the subsequent spatial aggregation step of spatiotemporally decoupled heterogeneous graph networks.

[0011] Optionally, the output space aggregation embedding specifically includes: All valid relation type subgraphs are input into the spatial aggregation unit in parallel. Graph convolution operation is performed independently for each subgraph. The normalized graph Laplacian operator is used to symmetrically normalize the adjacency matrix of the subgraph. The normalized adjacency matrix is ​​multiplied by the node feature matrix and then input into two graph convolution layers to obtain the subgraph-level embedding. The spatial aggregation embedding at the superedge level is obtained by weighting and summing each subgraph-level embedding according to the ratio of the number of nodes in the subgraph to the total number of nodes in the superedge. All hyperedges are spatially aggregated and embedded and stacked into a three-dimensional tensor in the order of their appearance. The first dimension of the three-dimensional tensor corresponds to the hyperedge index, the second dimension corresponds to different times within the time window, and the third dimension corresponds to the embedding vector dimension. The 3D tensor is input into the graph-level attention pooling layer. The pooling layer calculates the global context vector for each hyperedge embedding. The dot product of the global context vector and the hyperedge embedding is used as the attention score. After normalizing the attention score, the hyperedge embeddings are weighted and summed to output the global space aggregated embedding.

[0012] Optionally, the output time aggregation embedding includes the following steps: The global space aggregation embedding is divided into several time steps along the time dimension, and each time step corresponds to a preset time window; The embedded sequence is input into the gated recurrent unit network, which contains two hidden layers. The number of nodes in each hidden layer is twice the dimension of the embedded vector. The first hidden layer receives the global spatial aggregation embedding of the current time step and the hidden state output from the previous time step. The second hidden layer performs a nonlinear transformation on the output of the first layer. The gated recurrent unit network outputs a hidden state vector at each time step, and the hidden state vector of the last time step is used as the temporal aggregation embedding output. Temporal attention weights are introduced during the training phase of the gated recurrent unit network. The temporal attention weights are calculated by the dot product of the hidden state vector and the learnable temporal query vector at each time step. After normalizing the temporal attention weights, the hidden state vectors of all time steps are weighted and summed. The weighted sum is then concatenated with the hidden state vector of the last time step to obtain the temporal aggregation embedding output.

[0013] Optionally, the topology diagram of the output abnormal asset cluster specifically includes: Asset nodes with health scores below the warning threshold are marked as abnormal nodes. The minimum connected subgraph containing all abnormal nodes is extracted from the knowledge graph. The original hyperedges between abnormal nodes and the connection hyperedges between abnormal nodes and healthy nodes are preserved in the minimum connected subgraph. For each hyperedge in the minimum connected subgraph, calculate the abnormal density. The abnormal density is equal to the ratio of the number of abnormal nodes in the hyperedge to the total number of nodes in the hyperedge. Delete the hyperedges with abnormal density lower than the preset density threshold to obtain the abnormal association subgraph. The abnormal association subgraph is input into the graph layout algorithm. The graph layout algorithm uses a force-directed model to calculate the coordinates of each abnormal node on the two-dimensional plane. The attraction coefficient in the force-directed model is positively correlated with the historical influence strength between nodes inside the hyperedge, and the repulsion coefficient is positively correlated with the square root of the node degree. Nodes are color-coded based on their health scores, with darker colors corresponding to lower health scores. Edges are width-coded based on their anomaly density in the anomaly association subgraph, with wider edges corresponding to higher anomaly density. The output is a topology graph with color and line width annotations.

[0014] The beneficial effects of this invention are: First, we use temporal causality testing to dynamically prune superedges, removing redundant dependencies with no significant causal impact. This reduces the size of the asset knowledge graph by more than 40%, shrinks the search space for influence propagation paths, and improves computational efficiency.

[0015] Second, a multi-head attention decomposition module is introduced to split multi-dependency hyperedges into eight types of relation subgraphs, which respectively model physical dependencies, logical derivation, temporal triggering, governance constraints, value transmission, redundancy backup, version inheritance and access authorization semantics, to achieve accurate capture of implicit relationships across systems and reduce the false negative rate.

[0016] Third, spatial aggregation and temporal aggregation are processed separately through spatiotemporal decoupling heterogeneous graph networks. Spatial aggregation units aggregate neighbor features along the hyperedge structure, while temporal aggregation units use gated cyclic units to capture temporal evolution patterns, outputting asset health scores and risk warning signals. The topology diagram of abnormal asset clusters intuitively presents the risk transmission path, supporting maintenance personnel to quickly locate the root cause. Attached Figure Description

[0017] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is a flowchart of a knowledge graph-based enterprise digital asset management method proposed in this invention. Figure 2 This is a schematic diagram of temporal causal superedge pruning for an enterprise digital asset management method based on knowledge graph proposed in this invention; Figure 3 This is a schematic diagram of multi-head attention hyperedge decomposition for an enterprise digital asset management method based on knowledge graphs proposed in this invention. Detailed Implementation

[0018] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0019] refer to Figure 1-3 A knowledge graph-based enterprise digital asset management method includes the following steps: Acquire multi-source heterogeneous data of enterprise digital assets, construct an initial asset knowledge graph, and use asset entities as nodes and multi-dimensional dependencies between assets as hyperedges; The causal effect strength of each hyperedge to key asset nodes is calculated by the temporal causal test method. Hyperedges with causal effect strength below a preset threshold are deleted to obtain a sparse hypergraph. Each hyperedge in the sparse hypergraph is input into the multi-head attention decomposition module. Based on the attention weights of the nodes inside the hyperedge for different relation types, each hyperedge is split into several relation type subgraphs, and each relation type subgraph corresponds to an asset relation semantic. Several relation type subgraphs are input in parallel into the spatial aggregation unit of the spatiotemporally decoupled heterogeneous graph network. The spatial aggregation unit aggregates the features of neighbor nodes along the hyperedge structure and outputs the spatial aggregation embedding. Spatial aggregation embedding is input into the temporal aggregation unit of the spatiotemporally decoupled heterogeneous graph network. The temporal aggregation unit aggregates the embedding sequence of several historical moments along the time axis and outputs the temporal aggregation embedding. By embedding time aggregation into the fully connected classification layer, a health score and risk warning signal for each asset node can be obtained. Mark the asset nodes whose health scores are below the warning threshold and their associated superedges as abnormal asset clusters, and output the topology diagram of the abnormal asset clusters.

[0020] In this embodiment, the multi-source heterogeneous data of enterprise digital assets specifically includes: Asset configuration records, asset metadata definitions, descriptions of dependencies between assets, asset change history logs, asset usage statistics, technical documentation text associated with assets, source code structure information corresponding to assets, and interface definition data provided by assets to the outside world.

[0021] In this embodiment, obtaining the sparsified hypergraph includes the following steps: Collect the time-series feature vector of each asset node within a continuous time window. The time-series feature vector includes asset change frequency, asset call response time, asset error rate, and the rate of change of the number of upstream and downstream connections of the asset. Select a set of critical asset nodes, which includes asset nodes that carry core business processes, convergence nodes that aggregate more than five dependent paths, and asset nodes that have appeared three or more times in historical fault records. For each hyperedge, extract the temporal feature vector sequence of all nodes within the hyperedge within the time window, input the temporal feature vector sequence into the structural vector autoregressive model, and estimate the conditional Granger causality test statistic of each node within the hyperedge to the key asset node. Set a causal strength threshold and define the causal strength of each hyperedge as the weighted average of the Granger causality test statistics of all nodes within the hyperedge to the key asset node. The weights are determined by the inverse of the graph distance between the nodes within the hyperedge and the key asset node. Remove hyperedges with causal strength below the causal strength threshold to obtain a preliminary sparsified hypergraph; The temporal feature vectors of the remaining hyperedges in the initial sparsified hypergraph are collected, and the historical causal evolution trajectory of each remaining hyperedge is constructed. The causal evolution trajectory contains the causal intensity sequence of three time windows. The causal evolution trajectory is input into the long short-term memory network to predict the causal intensity decay coefficient of each remaining superedge in the next time window; Set an upper limit for the decay coefficient. When the predicted decay coefficient is lower than the upper limit, retain the hyperedge. When the predicted decay coefficient is not lower than the upper limit, delete the hyperedge to obtain the final sparse hypergraph.

[0022] Specifically, the process of inputting the time-series feature vector sequence into the structure vector autoregressive model to estimate the conditional Granger causality test statistic for each node within the hyperedge with respect to the key asset node includes: For each hyperedge, the temporal feature vector of each asset node it contains is collected within a continuous time window. The feature dimensions include change frequency, response time, error rate and connection number change rate. Assets that carry core business processes or have frequent historical failures are selected as key nodes. For each key node, a structure vector autoregressive model is constructed, which takes the time series data of all nodes inside the hyperedge as input to predict the state of the key node. Conditional Granger causality tests were used to compare the changes in prediction error when using all nodes versus removing a certain node, and a statistic was calculated. A higher statistic indicates a stronger causal effect of that node on key nodes. The statistic of each node within the hyperedge was weighted and averaged according to the inverse of the graph distance to obtain a comprehensive causal strength score for the hyperedge, which was used for subsequent pruning decisions.

[0023] The causal evolution trajectory includes three time windows in the causal intensity sequence: The first time window is the historical baseline window, which is used to calculate the baseline of causal strength of the superedge in an earlier period, reflecting the causal dependency pattern between assets under stable operating conditions.

[0024] The second time window is the recent evolution window, which is used to calculate the change in causal strength of the superedge in the middle period and capture the recent fluctuation trend of causal relationship between assets.

[0025] The third time window is the current monitoring window, used to calculate the current value of the causal strength of the hyperedge in the most recent period, reflecting the latest state of causal dependence between assets.

[0026] Three windows of equal length, each set to five consecutive calendar days, are used, with no gaps or overlaps between adjacent windows. Using the current time as a baseline, the first window corresponds to days 16 to 20, the second to days 11 to 15, and the third to days 6 to 10. The causal strengths of the three windows are concatenated chronologically and input into a Long Short-Term Memory (LSTM) network to predict the decay trend of causal strength in the next time period (days 21 to 25). This three-segment window design takes into account the long-term baseline, medium-term trend, and recent state, making the decay coefficient prediction more accurate.

[0027] In this embodiment, splitting each hyperedge into several relation type subgraphs includes the following steps: For each hyperedge, construct a node feature matrix and an edge type encoding vector. The rows of the node feature matrix correspond to the embedding vector of each asset node within the hyperedge, and the columns correspond to the embedding dimension. The length of the edge type encoding vector is equal to the total number of predefined relation types. Each position takes a value of 0 or 1, where 1 indicates that this hyperedge contains a priori indication of the corresponding relation type. The node feature matrix and edge type encoding vector are input into the multi-head attention decomposition module. The multi-head attention decomposition module contains several parallel attention heads. Each attention head independently calculates a set of attention weight matrices. The elements of the attention weight matrix represent the association strength between nodes within the hyperedge under a specific relation type. Among the several parallel attention heads, some attention heads calculate the structural similarity weight by multiplying the transpose of the node feature matrix, another part of the attention heads calculate the semantic compatibility weight by multiplying the node feature matrix and the learnable type embedding matrix, and the remaining attention heads calculate the evolutionary consistency weight by multiplying the temporal difference features of the node feature matrix and the topological centrality weight by multiplying the degree vector of the node feature matrix, respectively. After summing the element-wise output weight matrices of all attention heads, the result is input into the softmax function to obtain the normalized relation assignment probability matrix. Each column of the relation assignment probability matrix corresponds to a predefined relation type, and each row corresponds to a node inside the hyperedge. For each predefined relation type, extract the node pairs whose probability values ​​in the column of that type in the relation assignment probability matrix exceed the assignment threshold, use these node pairs and their corresponding probability values ​​as edge weights, and construct the subgraph corresponding to this relation type. The subgraph only contains nodes whose probability values ​​in the nodes inside the hyperedge exceed the assignment threshold. Delete subgraphs with fewer than 2 nodes, and retain subgraphs with 2 or more nodes as valid relation type subgraphs.

[0028] Specifically, the calculation of structural similarity weights includes: For a hyperedge containing multiple asset nodes, first construct the node feature matrix of the hyperedge. The rows of the matrix correspond to each asset node in the hyperedge, and the columns correspond to the embedding vector dimension of each node. The node embedding vector expresses the adjacency structure information of the asset in the graph. To calculate the structural similarity weight between any two nodes within the hyperedge, a matrix multiplication operation is performed between the node's feature matrix and its transpose. The transpose is equivalent to swapping the rows and columns of the original matrix; that is, rows become columns, and columns become rows. Multiplying the original matrix by the transpose yields a square matrix with the number of rows and columns equal to the number of nodes within the hyperedge. The value of an element in a specific row and column of this square matrix represents the original structural similarity score between the corresponding two nodes within the hyperedge. The essence of this multiplication operation is to calculate the dot product of the embedding vectors of two nodes. The larger the dot product value, the closer the embedding vectors of the two nodes are, that is, the more similar the neighborhood structure of the two assets in the graph. For example, in a hyperedge containing database nodes, application interface nodes, and report nodes, the dot product value of the application interface node and the database node is higher, indicating that they are close in dependency structure; while the dot product value of the report node and the database node is lower, indicating that they have greater structural differences. After obtaining the original similarity matrix, each row of the matrix is ​​normalized so that the sum of the probability values ​​of all columns in the same row is a fixed value, thus obtaining the final structural similarity weight matrix. Each element of the weight matrix has a value between zero and one, representing the attention weight of one node to another when evaluated only from the structural similarity dimension. This weight will be fused with the semantic compatibility weight, evolutionary consistency weight, and topological centrality weight calculated by other attention heads to jointly determine the probability of the distribution of the relationship type between nodes within the hyperedge.

[0029] Specifically, calculating semantic compatibility weights includes: For a hyperedge containing multiple asset nodes, the original feature vector of each asset node and a learnable type embedding matrix are first obtained. Each row of the type embedding matrix corresponds to a predefined asset relationship semantic, such as physical dependency, logical derivation, and temporal triggering. The parameters of this matrix are automatically learned during model training, and each row represents the standard expression of a relationship semantic in the embedding space. To calculate the compatibility between any two nodes within a hyperedge under a specific relational semantics, the original feature vector of each node is multiplied by the type embedding matrix. Specifically, for each node within the hyperedge, its feature vector is multiplied by the type embedding matrix to obtain a vector. Each component of this vector represents the degree of matching between the node and the corresponding relational semantics. For example, for the "database server" node, its matching value is high in physical dependency semantics and low in version inheritance semantics. After obtaining the matching vectors of each node with various relational semantics, it is necessary to evaluate the semantic compatibility between two nodes. The calculation method is to perform a dot product operation on the matching vector of node A and the matching vector of node B. The larger the dot product value, the more consistent the expression of the two nodes in the relational semantic space, that is, the more likely there is a corresponding semantic relationship between them. For example, the dot product value of the database node and the application interface node in terms of physical dependency semantics is high, indicating that there is a high probability that the two have a physical dependency relationship; while the dot product value of the database node and the report node in terms of physical dependency semantics is low, indicating that the two are unlikely to form a physical dependency relationship. The above operation is repeated for all node pairs within the hyperedge to obtain a semantic compatibility score matrix, where the number of rows and columns equals the number of nodes within the hyperedge. Then, each row of this matrix is ​​normalized so that the sum of the probability values ​​of all columns in the same row is a fixed value, resulting in the final semantic compatibility weight matrix. Each element of this weight matrix has a value between zero and one, representing the attention weight of one node to another when evaluated solely from the semantic compatibility dimension, reflecting the degree of matching between the two assets at the semantic level of their relationship. This weight is then fused with structural similarity weights, evolutionary consistency weights, and topological centrality weights calculated by other attention heads to jointly determine the probability of assigning relationship types between nodes within the hyperedge.

[0030] The calculation of topological centrality weights specifically includes: Obtain the degree value of each node in the entire asset knowledge graph, including total degree, in-degree, out-degree, and eigenvector centrality score. Eigenvector centrality takes into account the importance of neighboring nodes, with a node frequently relied upon by core assets scoring higher. For each pair of nodes within the hyperedge, perform a dot product operation on the degree vectors of the two nodes. The larger the dot product value, the more similar the roles of the two nodes in the network topology. For example, the dot product value is high between two core data tables that are called by multiple upper-layer applications, while the dot product value is low between the core table and the edge log file. After calculating the scores for all node pairs, a score matrix is ​​obtained. Each row of the matrix is ​​normalized to obtain the topological centrality weight matrix, with element values ​​between zero and one, representing the attention weights between nodes evaluated from the topological centrality dimension. This weight is fused with the results of other attention heads to identify similar centrality asset pairs, such as servers that are backups of each other.

[0031] In this embodiment, the asset relationship semantics corresponding to each relationship type subgraph are determined by a predefined set of relationship types, which includes the following eight asset relationship semantics: Physical dependency semantics indicate that the operation of one asset instance requires computing or storage resources provided by another asset instance; Logical derivation semantics indicates that the data content or functional definition of one asset originates from another asset; The semantics of time-triggered relationships indicate that a change event of one asset will trigger another asset to perform a preset operation within a specified time window; Governance constraint semantics indicate that one asset must comply with the security policies or compliance rules defined by another asset. The semantics of value transmission relationships indicate that changes in the effectiveness of one asset will propagate along the graph and affect the business contribution of another asset; Redundancy and backup relationship semantics indicate that two assets are functionally interchangeable, and the other takes over in the event of a single point of failure; Version inheritance semantics indicate that one asset is a historical version or iterative derivative version of another asset; Access authorization relationship semantics, indicating that one asset holds the authorization credentials to read or write to another asset; The multi-head attention decomposition module outputs eight sets of relation type subgraphs for each hyperedge. Each set of subgraphs corresponds to one of the eight asset relation semantics mentioned above. The node pairs retained in the subgraphs represent asset node pairs within this hyperedge whose association probability value under this relation semantic exceeds a preset allocation threshold. When a superedge is decomposed by multi-head attention, if the number of nodes in three or more of the eight subgraphs is zero, this superedge is marked as a noisy superedge and deleted, and will not participate in the subsequent spatial aggregation step of spatiotemporally decoupled heterogeneous graph networks.

[0032] Among them, the semantics of time-triggered relationships indicate that a change event of one asset will trigger another asset to perform a preset operation within a specified time window, specifically including: The types and parameters of preset operations can be configured during system initialization, or can be dynamically added or modified during use. Each preset operation includes four attributes: operation type identifier, target asset location, execution condition judgment, and execution timeout. The execution condition judgment is used to specify that the operation is triggered only when specific conditions are met, such as data synchronization only being performed within a certain period of time on a working day. The execution timeout is used to specify the maximum waiting time for the operation. If the timeout is exceeded, the operation is judged to have failed and an exception log is recorded. By defining the semantics of the time-series trigger relationship, this invention can explicitly model the change response mechanism between assets into a knowledge graph, realizing the transformation from passive response to active scheduling.

[0033] In this embodiment, the output space aggregation embedding specifically includes: All valid relation type subgraphs are input into the spatial aggregation unit in parallel. Graph convolution operation is performed independently for each subgraph. The normalized graph Laplacian operator is used to symmetrically normalize the adjacency matrix of the subgraph. The normalized adjacency matrix is ​​multiplied by the node feature matrix and then input into two graph convolution layers to obtain the subgraph-level embedding. The spatial aggregation embedding at the superedge level is obtained by weighting and summing each subgraph-level embedding according to the ratio of the number of nodes in the subgraph to the total number of nodes in the superedge. All hyperedges are spatially aggregated and embedded and stacked into a three-dimensional tensor in the order of their appearance. The first dimension of the three-dimensional tensor corresponds to the hyperedge index, the second dimension corresponds to different times within the time window, and the third dimension corresponds to the embedding vector dimension. The 3D tensor is input into the graph-level attention pooling layer. The pooling layer calculates the global context vector for each hyperedge embedding. The dot product of the global context vector and the hyperedge embedding is used as the attention score. After normalizing the attention score, the hyperedge embeddings are weighted and summed to output the global space aggregated embedding.

[0034] Specifically, the parallel input space aggregation unit for all valid relation type subgraphs is: Simultaneously, multiple subgraphs split from a hyperedge are fed into different computational paths of the spatial aggregation unit. Each subgraph performs graph convolution operations independently, without waiting for each other. Specifically, the spatial aggregation unit pre-allocates a graph convolution channel for each relation type. For physically dependent subgraphs, channel one extracts the adjacency structure and node features of the subgraph, performs normalized convolution operations, and outputs the node embeddings from the perspective of that relation. For logically derived subgraphs, channel two performs the same convolution process; and so on for other subgraphs. The computation of each channel starts at the same time and runs in parallel, making full use of the parallel computing capabilities of the graphics processor. After all channels are completed, the embeddings output by each subgraph are weighted and summed according to the number of nodes in the subgraph, and then merged into a spatial aggregation embedding at the hyperedge level. This design avoids the linear delay caused by serial processing, and aggregation can be completed in one forward propagation regardless of the number of subgraphs.

[0035] Specifically, the symmetric normalization of the adjacency matrix of the subgraph using the normalized graph Laplacian operator includes: Each relation type subgraph is itself a small graph structure containing several asset nodes and connecting edges between nodes according to the relation semantics. To perform graph convolution operations on such a subgraph, the adjacency matrix of the subgraph must first be constructed. The adjacency matrix is ​​a square matrix with the number of rows and columns equal to the number of nodes in the subgraph. If there is an edge between the i-th node and the j-th node in the subgraph, the element in the i-th row and j-th column of the adjacency matrix is ​​1; otherwise, it is 0. For example, a physical dependency subgraph derived from a hyperedge contains three nodes: a database server, an application programming interface (API), and a report. There is a physical dependency edge between the database server and the API, and also between the API and the report. However, there is no direct edge between the database server and the report. The adjacency matrix of this subgraph is a 3x3 matrix, with the values ​​in the first row, second column, and second row, first column being 1, the values ​​in the second row, third column, and third row, second column being 1, and the remaining positions being 0. After obtaining the adjacency matrix, the normalized graph Laplacian operator can be used to perform symmetric normalization on it. Specifically, each element of the adjacency matrix is ​​divided by the square root of the product of the corresponding row degree and column degree, thereby eliminating the influence of node degree differences on the convolution result. The normalized adjacency matrix is ​​then multiplied by the node feature matrix to achieve the aggregation of neighbor features.

[0036] In this embodiment, the output time aggregation embedding includes the following steps: The global space aggregation embedding is divided into several time steps along the time dimension, and each time step corresponds to a preset time window; The embedded sequence is input into the gated recurrent unit network, which contains two hidden layers. The number of nodes in each hidden layer is twice the dimension of the embedded vector. The first hidden layer receives the global spatial aggregation embedding of the current time step and the hidden state output from the previous time step. The second hidden layer performs a nonlinear transformation on the output of the first layer. The gated recurrent unit network outputs a hidden state vector at each time step, and the hidden state vector of the last time step is used as the temporal aggregation embedding output. Temporal attention weights are introduced during the training phase of the gated recurrent unit network. The temporal attention weights are calculated by the dot product of the hidden state vector and the learnable temporal query vector at each time step. After normalizing the temporal attention weights, the hidden state vectors of all time steps are weighted and summed. The weighted sum is then concatenated with the hidden state vector of the last time step to obtain the temporal aggregation embedding output.

[0037] Specifically, the embedding sequence that divides the global spatial aggregation embedding into several time steps along the time dimension includes: According to a pre-set time window length and sliding step size, the continuously collected asset status data is divided into several non-overlapping or partially overlapping time periods. Each time period corresponds to a time step. For example, if the time window length is set to five days and the sliding step size is one day, starting from day zero, the first time step covers the data from day zero to day five, the second time step covers the data from day one to day six, and so on, generating a total of thirty-six time steps. Within each time step, the global spatial aggregation of all historical moments belonging to that time period is embedded and stacked in chronological order into a subsequence. The length of these subsequences is equal to the number of days within the time window, and the embedding vector dimension in each subsequence remains unchanged. These subsequences are then arranged in chronological order to form a three-dimensional array. The first dimension of the array is the time step index, the second dimension is the time index within the window, and the third dimension is the embedding vector dimension. This three-dimensional array serves as the input to the time aggregation unit, enabling the gated recurrent unit network to process the sequence information within each window sequentially according to the time step, thereby capturing the evolution pattern of the asset state on the time axis.

[0038] Specifically, the second hidden layer performs a nonlinear transformation on the output of the first layer, including: The hidden state vector output by the first hidden layer at each time step is used as the input of the second hidden layer. The second hidden layer first performs a linear mapping on the input, that is, multiplies the input vector with the weight matrix of the second layer and adds a bias term to obtain an intermediate vector. Then, a non-linear activation function is applied to each element of the intermediate vector. In this embodiment, the hyperbolic tangent function is used as the activation function to compress the input value to the range of negative one to positive one, thereby enhancing the expressive power of the model. After the nonlinear transformation, the second hidden layer outputs a new hidden state vector with a dimension equal to the number of nodes in this layer. This output serves two purposes: firstly, it is passed to the subsequent fully connected classification layer as the final hidden state at the current time step; secondly, it is retained for recursive computation in the next time step. The introduction of the nonlinear transformation enables the network to learn complex nonlinear patterns in asset state changes, such as the nonlinear decay trend of asset health scores over time.

[0039] Specifically, the introduction of temporal attention weights during the training phase of the gated recurrent unit network includes: During training, a set of weight parameters are automatically learned. These parameters are not fixed values ​​set in advance, but are gradually optimized from the training data through the backpropagation algorithm. Specifically, a random attention weight value is first initialized for each time step, which fluctuates in a small range near zero. When the training sample is input into the network, the network calculates the prediction result under the current attention weight and compares it with the true label to obtain the prediction error. Based on this error, the optimization algorithm calculates the gradient of the contribution of each attention weight to the error, and then fine-tunes the weight value along the gradient descent direction to reduce the error of the next prediction. Repeating the above process hundreds or thousands of times, the attention weights at each time step gradually converge to the optimal value. For example, in the asset health prediction task, after sufficient training, the weights of time steps closer to the current time are usually larger, while the weights of time steps farther from the current time are smaller, indicating that the network autonomously discovers that the recent state has a greater impact on the health score. Ultimately, the attention weight at each time step becomes a fixed value, which is used to perform a weighted summation of the hidden state vector.

[0040] In this embodiment, the output of the topology diagram of the abnormal asset cluster specifically includes: Asset nodes with health scores below the warning threshold are marked as abnormal nodes. The minimum connected subgraph containing all abnormal nodes is extracted from the knowledge graph. The original hyperedges between abnormal nodes and the connection hyperedges between abnormal nodes and healthy nodes are preserved in the minimum connected subgraph. For each hyperedge in the minimum connected subgraph, calculate the abnormal density. The abnormal density is equal to the ratio of the number of abnormal nodes in the hyperedge to the total number of nodes in the hyperedge. Delete the hyperedges with abnormal density lower than the preset density threshold to obtain the abnormal association subgraph. The abnormal association subgraph is input into the graph layout algorithm. The graph layout algorithm uses a force-directed model to calculate the coordinates of each abnormal node on the two-dimensional plane. The attraction coefficient in the force-directed model is positively correlated with the historical influence strength between nodes inside the hyperedge, and the repulsion coefficient is positively correlated with the square root of the node degree. Nodes are color-coded based on their health scores, with darker colors corresponding to lower health scores. Edges are width-coded based on their anomaly density in the anomaly association subgraph, with wider edges corresponding to higher anomaly density. The output is a topology graph with color and line width annotations.

[0041] Specifically, calculating the outlier density for each hyperedge in the minimum connected subgraph includes: The total number of all asset nodes contained within the superedge is recorded as the total number of nodes. The number of asset nodes marked as abnormal nodes within the superedge is counted and denoted as the number of abnormal nodes. An abnormal node is an asset node whose health score is lower than the warning threshold. Then, the anomaly density is calculated, which is the number of anomalous nodes divided by the total number of nodes, resulting in a ratio between zero and one. For example, if a hyperedge connects five asset nodes, and three of them have health scores below the warning threshold, then the anomaly density of that hyperedge is 0.6. As another example, if another hyperedge connects four asset nodes, and none of them are anomalous, then the anomaly density is zero. Set a preset density threshold, such as 0.5. When the abnormal density of a hyperedge is lower than this threshold, it is determined that the hyperedge is insufficient to represent the core propagation path of the abnormal asset cluster and is deleted from the minimum connected subgraph. When the abnormal density of a hyperedge is not lower than the threshold, the hyperedge is retained. In the example above, the hyperedge with an abnormal density of 0.6 is retained and the hyperedge with an abnormal density of zero is deleted. In this way, the resulting abnormal correlation subgraph only includes those superedges with a high proportion of abnormal nodes, thereby highlighting asset areas with highly concentrated risks and avoiding interference from edge dependence on root cause localization.

[0042] Example 1: To verify the feasibility and effectiveness of the knowledge graph-based enterprise digital asset management method proposed in this invention, it was applied to the IT asset management department of a large manufacturing enterprise. This enterprise possesses over 30,000 digital assets, covering database instances, application programming interfaces (APIs), data reports, configuration scripts, cloud service resources, middleware components, and various technical documents. These assets are distributed across eight different business departments, including R&D, production, sales, and finance, with complex dependencies between them.

[0043] In practice, the company faces two typical problems. First, when the underlying database needs upgrading, the operations and maintenance team cannot accurately predict which front-end reports and API interfaces will be affected. They often only receive complaints from business departments about the failures after the changes have been implemented, and on average, each database upgrade causes more than three business applications to malfunction. Second, asset health assessments rely on regular manual inspections. A health report containing two hundred core assets requires four engineers and five working days to complete, and the assessment results are highly subjective, with a miss rate as high as 30%.

[0044] To address the aforementioned issues, this invention employs a method for unified management of the enterprise's digital assets. First, raw data is extracted from the configuration management database, code repository, API gateway, process engine, and data catalog to construct an initial asset knowledge graph. In this graph, each asset entity is represented as a node, such as "Customer Information Table," "Order Processing API," "Inventory Synchronization Script," and "Daily Sales Report," totaling 28,400 nodes. The multi-factor dependencies between assets are modeled as hyperedges. A hyperedge can connect more than three asset nodes. For example, the "Order Processing API" depends on three assets: "Customer Information Table," "Product Catalog Table," and "Payment Gateway Configuration." These three dependencies are merged into a single hyperedge. The entire graph contains 12,300 hyperedges.

[0045] Next, a temporal causality test was performed. Twenty asset nodes carrying the core production process were selected as the critical asset set, including the "production scheduling system," "material requirements planning interface," and "quality traceability database." For each hyperedge, data was collected over a continuous 40-day time window, with a time step of one day. The temporal characteristics of each asset node included the number of changes, average response time, error rate, and connection number change rate. The temporal data was input into a structured vector autoregression model to calculate the Granger causality test statistic for each hyperedge to the critical assets. A causality strength threshold of 0.6 was set, and hyperedges with causality strength below this threshold were deleted. After pruning, 7,800 hyperedges were retained, representing a deletion rate of 36.6%. The deleted hyperedges were mainly intermittent temporary dependencies, such as a script connection used only for a single data migration.

[0046] The remaining 7,800 hyperedges are input into a multi-head attention decomposition module, which contains four parallel attention heads that calculate structural similarity weights, semantic compatibility weights, evolutionary consistency weights, and topological centrality weights, respectively. Each hyperedge is decomposed into eight relational subgraphs, corresponding to eight semantic types: physical dependency, logical derivation, temporal triggering, governance constraints, value transmission, redundancy backup, version inheritance, and access authorization. Taking the hyperedge containing the "inventory synchronization script" as an example, this hyperedge originally connected seven asset nodes. After decomposition, the connection to the database driver and storage device is retained in the physical dependency subgraph, the derivation relationship with the upstream order table and downstream inventory table is retained in the logical derivation subgraph, the triggering relationship with the daily midnight scheduling task is retained in the temporal triggering subgraph, and the association with the data security policy is retained in the governance constraint subgraph. The other four subgraphs have zero nodes in this hyperedge, and according to the rules, this hyperedge is marked as a noisy hyperedge and deleted. Ultimately, out of 12,300 superedges, 950 were identified as noisy superedges and deleted, leaving 11,350 valid superedges.

[0047] All relational subgraphs are input in parallel into the spatial aggregation unit of a spatiotemporally decoupled heterogeneous graph network. The spatial aggregation unit independently performs two layers of graph convolution operations on each subgraph, aggregating neighbor node features and outputting a 128-dimensional spatial aggregation embedding. This spatial aggregation embedding is then input into the temporal aggregation unit, which employs a gated recurrent unit network with two hidden layers. It aggregates historical embedding sequences along a 40-day timeline, outputting the temporal aggregation embedding. This temporal aggregation embedding is then input into a fully connected classification layer. This classification layer contains three output nodes, corresponding to high, medium, and low health scores, and uses a softmax function to output a probability distribution. The probability value of the lowest score is taken as the health risk score; an alert is triggered when this score exceeds 0.7.

[0048] After running the method of this invention for three months, the asset management effectiveness of the enterprise was evaluated. Meanwhile, a traditional asset management method based on a configuration management database and manual inspection was used as a control group and run for the same period. The evaluation data is shown in the table below.

[0049] Table 1: Performance Comparison of the Invention Method and Existing Methods in Digital Asset Management

[0050] The data in the table shows that the method of this invention outperforms traditional methods in multiple dimensions. The time required for asset change impact analysis has been reduced from 185 minutes to 23.5 minutes, improving processing efficiency by nearly eight times. Regarding implicit dependency detection, traditional methods are completely unable to automatically detect them, while the method of this invention discovers an average of 147 implicit dependencies per month. For example, in one analysis, it automatically identified that a field change in the "Quality Traceability Database" would affect the "After-Sales Analysis Report," which is located more than five steps away, and the person in charge of that report was previously unaware of this dependency. The accuracy rate of asset health assessment has increased from 68.5% to 94.3%, with a significant reduction in the false negative rate.

[0051] The response time for operational and maintenance (O&M) failures has been reduced from 7.8 hours to 1.9 hours, and risk warnings can be issued 28.6 hours in advance, giving the O&M team ample time for preventative maintenance. In one specific case, the method of this invention provided a warning 32 hours in advance that the health score of the "Material Requirements Planning (MRP) interface" would drop from 0.85 to 0.32 because the "Supplier Data Table," on which it depends, underwent four structural changes in the past five days. The O&M team intervened immediately after the warning, adjusting the interface's query logic and averting a potential production interruption. Traditional methods only receive complaints after a failure occurs, with an average recovery time of 7.8 hours.

[0052] Furthermore, the workload of manual inspections decreased from 160 person-hours per week to 36 person-hours, meaning that the work that originally required four full-time engineers can now be completed by one engineer within standard working hours. The number of production accidents caused by changes decreased from 12 per quarter to three, with the remaining three accidents occurring on peripheral assets not yet covered by the system, indicating that the method of this invention has a significant effect on the coverage of core assets. The semantic alignment time for cross-departmental assets decreased from 14 days to 2.5 days. Differences in the understanding of the same asset among different departments were unified through multi-head attention decomposition. For example, the "customer identifier" is called customerId in the R&D department, Client_Number in the sales department, and Cust_ID in the finance department. The method of this invention automatically identifies that these three nodes point to the same asset and establishes an equivalence relationship in the knowledge graph.

[0053] The aforementioned implementation data fully demonstrates the effectiveness of the method of this invention. By employing superedge temporal causal pruning and multi-head attention decomposition, combined with spatiotemporally decoupled heterogeneous graph networks, the method of this invention achieves significantly better technical results than traditional methods in terms of asset dependency discovery, accuracy of health assessment, timeliness of fault early warning, and operational efficiency.

[0054] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A knowledge graph-based enterprise digital asset management method, characterized in that, Includes the following steps: Acquire multi-source heterogeneous data of enterprise digital assets, construct an initial asset knowledge graph, and use asset entities as nodes and multi-dimensional dependencies between assets as hyperedges; The causal effect strength of each hyperedge to key asset nodes is calculated by the temporal causal test method. Hyperedges with causal effect strength below a preset threshold are deleted to obtain a sparse hypergraph. Each hyperedge in the sparse hypergraph is input into the multi-head attention decomposition module. Based on the attention weights of the nodes inside the hyperedge for different relation types, each hyperedge is split into several relation type subgraphs, and each relation type subgraph corresponds to an asset relation semantic. Several relation type subgraphs are input in parallel into the spatial aggregation unit of the spatiotemporally decoupled heterogeneous graph network. The spatial aggregation unit aggregates the features of neighbor nodes along the hyperedge structure and outputs the spatial aggregation embedding. Spatial aggregation embedding is input into the temporal aggregation unit of the spatiotemporally decoupled heterogeneous graph network. The temporal aggregation unit aggregates the embedding sequence of several historical moments along the time axis and outputs the temporal aggregation embedding. By embedding time aggregation into the fully connected classification layer, a health score and risk warning signal for each asset node can be obtained. Mark the asset nodes whose health scores are below the warning threshold and their associated superedges as abnormal asset clusters, and output the topology diagram of the abnormal asset clusters; The step of splitting each hyperedge into several relation-type subgraphs includes the following steps: For each hyperedge, construct a node feature matrix and an edge type encoding vector. The rows of the node feature matrix correspond to the embedding vector of each asset node within the hyperedge, and the columns correspond to the embedding dimension. The length of the edge type encoding vector is equal to the total number of predefined relation types. Each position takes a value of 0 or 1, where 1 indicates that this hyperedge contains a priori indication of the corresponding relation type. The node feature matrix and edge type encoding vector are input into the multi-head attention decomposition module. The multi-head attention decomposition module contains several parallel attention heads. Each attention head independently calculates a set of attention weight matrices. The elements of the attention weight matrix represent the association strength between nodes within the hyperedge under a specific relation type. Among the several parallel attention heads, some attention heads calculate the structural similarity weight by multiplying the transpose of the node feature matrix, another part of the attention heads calculate the semantic compatibility weight by multiplying the node feature matrix and the learnable type embedding matrix, and the remaining attention heads calculate the evolutionary consistency weight by multiplying the temporal difference features of the node feature matrix and the topological centrality weight by multiplying the degree vector of the node feature matrix, respectively. After summing the element-wise output weight matrices of all attention heads, the result is input into the softmax function to obtain the normalized relation assignment probability matrix. Each column of the relation assignment probability matrix corresponds to a predefined relation type, and each row corresponds to a node inside the hyperedge. For each predefined relation type, extract the node pairs whose probability values ​​in the column of that type in the relation assignment probability matrix exceed the assignment threshold, use these node pairs and their corresponding probability values ​​as edge weights, and construct the subgraph corresponding to this relation type. The subgraph only contains nodes whose probability values ​​in the nodes inside the hyperedge exceed the assignment threshold. Delete subgraphs with fewer than 2 nodes, and retain subgraphs with 2 or more nodes as valid relation type subgraphs.

2. The enterprise digital asset management method based on knowledge graphs according to claim 1, characterized in that, The multi-source heterogeneous data of the enterprise's digital assets specifically includes: Asset configuration records, asset metadata definitions, descriptions of dependencies between assets, asset change history logs, asset usage statistics, technical documentation text associated with assets, source code structure information corresponding to assets, and interface definition data provided by assets to the outside world.

3. The enterprise digital asset management method based on knowledge graphs according to claim 2, characterized in that, The process of obtaining the sparsified hypergraph includes the following steps: Collect the time-series feature vector of each asset node within a continuous time window. The time-series feature vector includes asset change frequency, asset call response time, asset error rate, and the rate of change of the number of upstream and downstream connections of the asset. Select a set of critical asset nodes, which includes asset nodes that carry core business processes, convergence nodes that aggregate more than five dependent paths, and asset nodes that have appeared three or more times in historical fault records. For each hyperedge, extract the temporal feature vector sequence of all nodes within the hyperedge within the time window, input the sequence into the structure vector autoregressive model, and estimate the conditional Granger causality test statistic of each node within the hyperedge to the key asset node. Set a causal strength threshold and define the causal strength of each hyperedge as the weighted average of the Granger causality test statistics of all nodes within the hyperedge to the key asset node. The weights are determined by the inverse of the graph distance between the nodes within the hyperedge and the key asset node. Remove hyperedges with causal strength below the causal strength threshold to obtain a preliminary sparsified hypergraph; The temporal feature vectors of the remaining hyperedges in the initial sparsified hypergraph are collected, and the historical causal evolution trajectory of each remaining hyperedge is constructed. The causal evolution trajectory contains the causal intensity sequence of three time windows. The causal evolution trajectory is input into the long short-term memory network to predict the causal intensity decay coefficient of each remaining superedge in the next time window; Set an upper limit for the decay coefficient. When the predicted decay coefficient is lower than the upper limit, retain the hyperedge. When the predicted decay coefficient is not lower than the upper limit, delete the hyperedge to obtain the final sparse hypergraph.

4. The enterprise digital asset management method based on knowledge graphs according to claim 1, characterized in that, The asset relationship semantics corresponding to each relationship type subgraph are determined by a predefined set of relationship types, which includes the following eight asset relationship semantics: Physical dependency semantics indicate that the operation of one asset instance requires computing or storage resources provided by another asset instance; Logical derivation semantics indicates that the data content or functional definition of one asset originates from another asset; The semantics of time-triggered relationships indicate that a change event of one asset will trigger another asset to perform a preset operation within a specified time window; Governance constraint semantics indicate that one asset must comply with the security policies or compliance rules defined by another asset. The semantics of value transmission relationships indicate that changes in the effectiveness of one asset will propagate along the graph and affect the business contribution of another asset; Redundancy and backup relationship semantics indicate that two assets are functionally interchangeable, and the other takes over in the event of a single point of failure; Version inheritance semantics indicate that one asset is a historical version or iterative derivative version of another asset; Access authorization relationship semantics, indicating that one asset holds the authorization credentials to read or write to another asset; The multi-head attention decomposition module outputs eight sets of relation type subgraphs for each hyperedge. Each set of subgraphs corresponds to one of the eight asset relation semantics mentioned above. The node pairs retained in the subgraphs represent asset node pairs within this hyperedge whose association probability value under this relation semantic exceeds a preset allocation threshold. When a superedge is decomposed by multi-head attention, if the number of nodes in three or more of the eight subgraphs is zero, this superedge is marked as a noisy superedge and deleted, and will not participate in the subsequent spatial aggregation step of spatiotemporally decoupled heterogeneous graph networks.

5. The enterprise digital asset management method based on knowledge graphs according to claim 4, characterized in that, The output space aggregation embedding specifically includes: All valid relation type subgraphs are input into the spatial aggregation unit in parallel. Graph convolution operation is performed independently for each subgraph. The normalized graph Laplacian operator is used to symmetrically normalize the adjacency matrix of the subgraph. The normalized adjacency matrix is ​​multiplied by the node feature matrix and then input into two graph convolution layers to obtain the subgraph-level embedding. The spatial aggregation embedding at the superedge level is obtained by weighting and summing each subgraph-level embedding according to the ratio of the number of nodes in the subgraph to the total number of nodes in the superedge. All hyperedges are spatially aggregated and embedded and stacked into a three-dimensional tensor in the order of their appearance. The first dimension of the three-dimensional tensor corresponds to the hyperedge index, the second dimension corresponds to different times within the time window, and the third dimension corresponds to the embedding vector dimension. The 3D tensor is input into the graph-level attention pooling layer. The pooling layer calculates the global context vector for each hyperedge embedding. The dot product of the global context vector and the hyperedge embedding is used as the attention score. After normalizing the attention score, the hyperedge embeddings are weighted and summed to output the global space aggregated embedding.

6. The enterprise digital asset management method based on knowledge graphs according to claim 5, characterized in that, The output time aggregation embedding includes the following steps: The global space aggregation embedding is divided into several time steps along the time dimension, and each time step corresponds to a preset time window; The embedded sequence is input into the gated recurrent unit network, which contains two hidden layers. The number of nodes in each hidden layer is twice the dimension of the embedded vector. The first hidden layer receives the global spatial aggregation embedding of the current time step and the hidden state output from the previous time step. The second hidden layer performs a nonlinear transformation on the output of the first layer. The gated recurrent unit network outputs a hidden state vector at each time step, and the hidden state vector of the last time step is used as the temporal aggregation embedding output. Temporal attention weights are introduced during the training phase of the gated recurrent unit network. The temporal attention weights are calculated by the dot product of the hidden state vector and the learnable temporal query vector at each time step. After normalizing the temporal attention weights, the hidden state vectors of all time steps are weighted and summed. The weighted sum is then concatenated with the hidden state vector of the last time step to obtain the temporal aggregation embedding output.

7. The enterprise digital asset management method based on knowledge graphs according to claim 6, characterized in that, The topology diagram of the output abnormal asset cluster specifically includes: Asset nodes with health scores below the warning threshold are marked as abnormal nodes. The minimum connected subgraph containing all abnormal nodes is extracted from the knowledge graph. The original hyperedges between abnormal nodes and the connection hyperedges between abnormal nodes and healthy nodes are preserved in the minimum connected subgraph. For each hyperedge in the minimum connected subgraph, calculate the abnormal density. The abnormal density is equal to the ratio of the number of abnormal nodes in the hyperedge to the total number of nodes in the hyperedge. Delete the hyperedges with abnormal density lower than the preset density threshold to obtain the abnormal association subgraph. The abnormal association subgraph is input into the graph layout algorithm. The graph layout algorithm uses a force-directed model to calculate the coordinates of each abnormal node on the two-dimensional plane. The attraction coefficient in the force-directed model is positively correlated with the historical influence strength between nodes inside the hyperedge, and the repulsion coefficient is positively correlated with the square root of the node degree. Nodes are color-coded based on their health scores, with darker colors corresponding to lower health scores. Edges are width-coded based on their anomaly density in the anomaly association subgraph, with wider edges corresponding to higher anomaly density. The output is a topology graph with color and line width annotations.

Images