A decision analysis method based on a full-factor time-varying knowledge graph

Abstract

The application discloses a kind of decision analysis methods based on full-element time-varying knowledge graph, it is related to atlas analysis technical field, including the following implementation steps: enter into hyperbolic capture end, by real-time acquisition data of different sources, and distinguish abnormal atlas;Enter into knowledge fusion end, by real-time to entity is semantically disambiguated, realize knowledge fusion and automatic identification sorting, enter entity identification end, by the GRU neural network layer and multi-head attention mechanism layer of two-way, first learn local information, then learn global information according to local information, realize distributed learning.The decision analysis method based on full-element time-varying knowledge graph, real-time extraction and remind abnormal in knowledge graph decision analysis, improve the timeliness and accuracy of knowledge graph, real-time prediction abnormal entity path, so that knowledge fusion is convenient to align complementary, optimize decision analysis to enhance the path search efficiency based on full-element time-varying knowledge graph.

Description

Technical Field

[0001] This invention relates to the field of graph analysis technology, and in particular to a decision analysis method based on a time-varying knowledge graph of all elements. Background Technology

[0002] Decision analysis based on full-factor time-varying knowledge graphs is a method that uses integrated information and data from multiple aspects of time dimension to support complex decision-making processes through analysis and reasoning. In today's data-driven era, traditional data analysis methods can no longer meet the increasingly complex decision-making needs. Knowledge graphs, as a structured way of representing knowledge, provide a richer and more dynamic form of data organization. In particular, after introducing time information, full-factor time-varying knowledge graphs can more accurately describe the dynamic changes of things in the real world, thereby providing strong support for decision-making.

[0003] Application number CN202310219566.2 discloses a decision analysis method and system based on big data. The technical solution adopted is as follows: This method involves sorting, classifying, mining, and refining aggregated full-volume enterprise data to establish analytical indicators and construct a data indicator system; it also constructs an analytical model to conduct multi-dimensional visual analysis of enterprises within a region, and applies it to self-service decision analysis for different users, providing various analytical indicators required for management decisions for personnel in different departments and positions, supporting various decision-making scenarios; specifically, it includes: enterprise data collection; decision data management; data indicator system construction; and self-service decision analysis.

[0004] A search of the aforementioned patents revealed several shortcomings in decision analysis based on full-element time-varying knowledge graphs: 1. Due to the rapid updates and changes in knowledge, knowledge graphs contain many complex hierarchical structures, making it relatively difficult to preserve such complex hierarchical structures in a linear space. Maintaining the timeliness and accuracy of knowledge graphs is a major challenge; 2. Historical data may conflict with new data, making timely semantic disambiguation impossible. Furthermore, knowledge fusion is not conducive to alignment and complementarity, and it cannot predict entity paths in the knowledge graph in real time; 3. The large differences in scale between multi-source knowledge graphs mean that capturing entities and concepts requires handling structures and search spaces of different scales, resulting in low efficiency and poor performance in decision analysis. Optimizing decision analysis to improve path search efficiency based on full-element time-varying knowledge graphs and effectively identifying decision labels are current challenges.

[0005] Therefore, a novel decision analysis method based on a full-factor time-varying knowledge graph is proposed to solve the above problems. Summary of the Invention

[0006] The main objective of this invention is to provide a decision analysis method based on a full-factor time-varying knowledge graph to solve the problems mentioned in the background above.

[0007] To achieve the above objectives, the technical solution adopted by the present invention is as follows: a decision analysis method based on a full-element time-varying knowledge graph, the method comprising the following implementation steps:

[0008] Step 1: Enter the hyperbolic capture terminal, collect data from different sources in real time, describe the same fact from different perspectives, form a heterogeneous knowledge graph, complement the facts in real time, and distinguish abnormal graphs;

[0009] Step 2: Enter the knowledge fusion terminal, perform semantic disambiguation on entities in real time, realize knowledge fusion and automatic identification and sorting, perform pattern alignment on entities that cannot complement each other, realize the mutual association between entities, extract abnormal entities, and predict the path of abnormal entities in real time.

[0010] Step 3: Enter the entity recognition end. Through the bidirectional GRU neural network layer and the multi-head attention mechanism layer, local information is first learned, and then global information is learned based on the local information to achieve distributed learning. Label prediction is performed based on the global information. The hidden state and the output of the attention mechanism are combined to achieve efficient entity recognition.

[0011] The hyperbolic capture end includes a hyperbolic receiving module, a hyperbolic complementary module, and a heterogeneous differentiation module;

[0012] The hyperbolic receiver module includes a multi-dimensional acquisition unit and a multi-dimensional description unit;

[0013] The multi-dimensional acquisition unit is used to acquire knowledge graphs from different angles in real time using a data acquisition instrument;

[0014] The multidimensional description unit is used to track and describe the same knowledge fact from different perspectives in real time using a data tracker.

[0015] The hyperbolic complementary module includes a hyperbolic construction unit and a knowledge complementarity unit;

[0016] The hyperbolic construction unit is used to construct a decision analysis model for time-varying knowledge graphs using a hyperbolic neural network;

[0017] The knowledge complementation unit is used to complement the differences in the knowledge graph of the decision analysis model, as follows:

[0018] Step 1: Calculate the affinity matrix in the decision analysis model at the current time step to measure the similarity of node attribute information. The formula is as follows:

[0019]

[0020] Where C represents the affinity matrix in the decision analysis model at the current moment, which measures the similarity of node attribute information, and x i and y iThis represents the i-th vector in the knowledge graph information and the knowledge graph information in the decision analysis model at the current moment;

[0021] Step 2: When two vectors are more similar, the closer their similarity value is to 1, the closer the knowledge graphs are; conversely, the closer they are, the greater the differences between the knowledge graphs.

[0022] The heterogeneous differentiation module includes a heterogeneous knowledge unit, an anomaly differentiation unit, and an anomaly alert unit;

[0023] The heterogeneous knowledge unit is used to identify one-quarter of the knowledge graphs with a similarity value less than or equal to 1 as abnormal knowledge.

[0024] The anomaly differentiation unit is used to perform threshold stratification based on the similarity of anomaly knowledge;

[0025] The anomaly alert unit is used to report the hierarchical knowledge graph of the anomaly knowledge to the system and issue a voice alarm.

[0026] The knowledge fusion terminal includes a semantic disambiguation module, a fusion ranking module, and a path prediction module;

[0027] The semantic disambiguation module is used to measure the degree of association between semantics and context in the knowledge graph by calculating the mutual information between them in real time, as follows:

[0028] Step 1: For each word meaning, calculate its mutual information with each word in the context;

[0029] Step II: Accumulate all mutual information values ​​to obtain the total score, and sort the semantics of the current time step by number according to the sequence number;

[0030] Step 3: Select the word meaning with the highest total score at the current moment as the final disambiguation result, and then proceed to the next semantic disambiguation step simultaneously.

[0031] The fusion sorting module includes a knowledge fusion unit and an identifier sorting unit;

[0032] The knowledge fusion unit is used to achieve real-time fusion of the knowledge graph by fusing the knowledge graph at the current moment, and to record it in real time by a data recorder.

[0033] The identification and sorting unit is used to color-code the knowledge graph after knowledge fusion and to arrange it by sequence number.

[0034] The path prediction module includes an alignment and association unit, an anomaly extraction unit, and a path prediction unit.

[0035] The alignment and association unit is used to align and associate knowledge graphs with high similarity in real time by utilizing the fact that the spatial capacity of hyperbolic space increases exponentially with the radius, and by using graph neural networks to capture the high-order proximity of entities through the neighborhood information of entities.

[0036] The anomaly extraction unit is used to automatically extract anomaly knowledge graphs that meet the threshold stratification when knowledge graph differences occur.

[0037] The path prediction unit is used to calculate the deviation angle of the abnormal knowledge graph in real time. The calculation formula is as follows:

[0038]

[0039] Where α is the angle between the path distance of the abnormal knowledge graph and the positive direction of the X-axis, θ is the angle between the movement distance of the abnormal graph and the positive direction of the Y-axis, A is the height of the abnormal knowledge graph, Δx is the vector of the abnormal knowledge graph in the X-axis direction, Δy is the vector of the abnormal knowledge graph in the Y-axis direction, and vector = 0 indicates that the abnormal knowledge graph path is normal, while vector ≠ 0 indicates that the abnormal knowledge graph path deviates.

[0040] The entity recognition module includes a two-layer learning module, a label prediction module, and an entity recognition module.

[0041] The two-layer learning module includes a bidirectional GRU neural network unit and a multi-source attention mechanism unit;

[0042] The bidirectional GRU neural network unit is used to learn local information in the knowledge graph through the bidirectional GRU neural network.

[0043] The multi-source attention mechanism unit is used to learn global information based on local information through BIGRU LAN.

[0044] The label prediction module includes a local progressive unit and a label prediction unit;

[0045] The local progressive unit is used to learn progressive global information through a local progressive mode;

[0046] The label prediction unit is used to predict labels based on the learned information.

[0047] The entity recognition module includes an output combining unit, an entity recognition unit, and a recognition verification unit;

[0048] The output combining unit is used to combine the outputs of the hidden state and the attention mechanism.

[0049] The entity recognition unit is used to perform entity recognition based on the label prediction results. The recognition method is as follows:

[0050] Step 1 (1): First, calculate the sum of the values ​​of the first layer and all candidate labels x. l =x1 l x2 l ,…,x|L| l An attention matrix α is generated through attention, which consists of the latent label distribution of each knowledge graph. Q = H is defined as follows. l K = V = x l ,in This uses a tag set representation, where |L| represents the total number of tags. Note that the output of matrix a is calculated as follows:

[0051] H l =attention(Q, K, V) = aV

[0052]

[0053] Step 1 (2): The label attention inference layer uses a multi-head attention mechanism. The calculation formula for multi-head attention is as follows:

[0054] H l = concat(head, ..., head) K )+H l

[0055] head i =attention(QW i Q KW i K VW i V )

[0056] Among them, W i Q W i K and W i V Here, k represents the parameters that need to be learned during the training process, and k is the number of threads.

[0057] Step 1 (3): The final output of the BIGRU LAN layer combines the hidden state of BIGRU and the output of the attention mechanism. Let H represent the final output, i.e., H = [H 1 H l ];

[0058] Step IV: The output of the first layer is then used as the input to the second BIGRULAN layer. The second BIGRULAN layer then predicts the labels of the knowledge graph based on the attention weights. The prediction formula is as follows:

[0059]

[0060] yi = argmax j (yi 1 , ..., yi n )

[0061] Among them, yi n Let represent the nth candidate label of the i-th knowledge graph, and yi represent the predicted label of the i-th knowledge graph in the sequence;

[0062] The identification verification unit is used to verify the entity recognition result in real time. It compares the recognition result with the knowledge graph set by the system. If there is a difference in the comparison, it indicates that the recognition result is invalid. The system is then notified to issue a voice reminder and repeat the entity recognition process. If there is no difference in the comparison, it indicates that the recognition result is valid.

[0063] The present invention has the following beneficial effects:

[0064] 1. In this invention, by setting up a hyperbolic capture end, in decision analysis based on a full-element time-varying knowledge graph, data from different sources are collected in real time and facts are complemented, so that different knowledge graphs can be complemented in fact during decision analysis, and automatic complementation is achieved through the difference ratio. Furthermore, by using heterogeneous knowledge and distinguishing abnormal graphs, anomalies in knowledge graph decision analysis are extracted and alerted in real time, thereby improving the timeliness and accuracy of the knowledge graph.

[0065] 2. In this invention, by setting up a knowledge fusion terminal, in decision analysis based on a full-element time-varying knowledge graph, knowledge fusion is achieved and automatically identified and sorted by performing semantic disambiguation on entities in real time. For entities that cannot complement each other, pattern alignment is performed, which can perform semantic disambiguation in a timely manner. At the same time, through the mutual association between entities, abnormal entity paths are predicted in real time, making knowledge fusion easier to align and complement each other, and enabling the prediction of entity paths in the knowledge graph in advance.

[0066] 3. In this invention, by setting up an entity recognition end, in decision analysis based on a full-element time-varying knowledge graph, distributed learning is achieved through a bidirectional GRU neural network layer and a multi-head attention mechanism layer, reducing the scale difference between multi-source knowledge graphs, and label prediction is performed based on global information. The combination of hidden state and attention mechanism output achieves efficient entity recognition, which can optimize decision analysis to improve the path search efficiency based on a full-element time-varying knowledge graph. Effective identification of decision labels is the current stage. Attached Figure Description

[0067] Figure 1 This is an overall flowchart of a decision analysis method based on a full-element time-varying knowledge graph according to the present invention;

[0068] Figure 2This is a schematic diagram of the architecture of a decision analysis method based on a full-element time-varying knowledge graph according to the present invention;

[0069] Figure 3 This is a schematic diagram of the architecture of a decision analysis method based on a full-element time-varying knowledge graph according to the present invention;

[0070] Figure 4 This is a schematic diagram of the architecture of a decision analysis method based on a full-element time-varying knowledge graph according to the present invention. Detailed Implementation

[0071] To make the technical means, creative features, objectives and effects of this invention easier to understand, the invention will be further described below in conjunction with specific embodiments.

[0072] Example 1

[0073] Please refer to Figure 1 and Figure 2 As shown: A decision analysis method based on a full-factor time-varying knowledge graph includes the following implementation steps:

[0074] Step 1: Enter the hyperbolic capture terminal, collect data from different sources in real time, describe the same fact from different perspectives, form a heterogeneous knowledge graph, complement the facts in real time, and distinguish abnormal graphs;

[0075] Step 2: Enter the knowledge fusion terminal, perform semantic disambiguation on entities in real time, realize knowledge fusion and automatic identification and sorting, perform pattern alignment on entities that cannot complement each other, realize the mutual association between entities, extract abnormal entities, and predict the path of abnormal entities in real time.

[0076] Step 3: Enter the entity recognition end. Through the bidirectional GRU neural network layer and the multi-head attention mechanism layer, local information is first learned, and then global information is learned based on the local information to achieve distributed learning. Label prediction is performed based on the global information. The hidden state and the output of the attention mechanism are combined to achieve efficient entity recognition.

[0077] The hyperbolic capture end includes a hyperbolic receiving module, a hyperbolic complementary module, and a heterogeneous differentiation module;

[0078] The hyperbolic receiver module includes a multidimensional acquisition unit and a multidimensional description unit;

[0079] The multi-dimensional acquisition unit is used to acquire knowledge graphs from different perspectives in real time using a data acquisition instrument;

[0080] Multidimensional description units are used to track and describe the same knowledge fact from different perspectives in real time using a data tracker.

[0081] The hyperbolic complementary module includes hyperbolic construction units and knowledge complementarity units;

[0082] Hyperbolic building blocks are used to construct decision analysis models for time-varying knowledge graphs using hyperbolic neural networks;

[0083] The knowledge complementarity unit is used to complement the differences in the knowledge graphs within the decision analysis model, as follows:

[0084] Step 1: Calculate the affinity matrix in the decision analysis model at the current time step to measure the similarity of node attribute information. The formula is as follows:

[0085]

[0086] Where C represents the affinity matrix in the decision analysis model at the current moment, which measures the similarity of node attribute information, and x i and y i This represents the i-th vector in the knowledge graph information and the knowledge graph information in the decision analysis model at the current moment;

[0087] Step 2: When two vectors are more similar, the closer their similarity value is to 1, the closer the knowledge graphs are; conversely, the closer they are, the greater the differences between the knowledge graphs.

[0088] The heterogeneity differentiation module includes heterogeneous knowledge units, anomaly differentiation units, and anomaly alert units;

[0089] Heterogeneous knowledge units are used to identify one-quarter of knowledge graphs with a similarity value less than or equal to 1 as anomalous knowledge.

[0090] The anomaly differentiation unit is used to perform threshold stratification based on the similarity of anomaly knowledge;

[0091] The anomaly alert unit is used to report the hierarchical knowledge graph of anomalies to the system and issue a voice alarm.

[0092] By collecting data from different sources in real time and describing the same fact from different perspectives, heterogeneous knowledge graphs are formed and facts are complemented. This enables the different knowledge graphs to complement each other during decision analysis, ensuring that complex hierarchical structures can be preserved in the linear space in real time and achieving automatic complementation. Furthermore, by identifying heterogeneous knowledge and distinguishing abnormal graphs, anomalies in knowledge graph decision analysis can be extracted and alerted in real time, improving the timeliness and accuracy of the knowledge graph.

[0093] Example 2

[0094] Please refer to Figure 3 As shown: Based on Embodiment 1, the knowledge fusion terminal includes a semantic disambiguation module, a fusion ranking module, and a path prediction module;

[0095] The semantic disambiguation module measures the degree of association between semantics and context in a knowledge graph by calculating the mutual information between them in real time, as follows:

[0096] Step 1: For each word meaning, calculate its mutual information with each word in the context;

[0097] Step II: Accumulate all mutual information values ​​to obtain the total score, and sort the semantics of the current time step by number according to the sequence number;

[0098] Step 3: Select the word meaning with the highest total score at the current moment as the final disambiguation result, and then proceed to the next semantic disambiguation step simultaneously.

[0099] The fusion sorting module includes a knowledge fusion unit and an identifier sorting unit;

[0100] The knowledge fusion unit is used to achieve real-time fusion of the knowledge graph by fusing the knowledge graph at the current moment, and records the data in real time through a data logger.

[0101] The identification and sorting unit is used to color-code and number the knowledge graph after knowledge fusion.

[0102] The path prediction module includes an alignment and association unit, an anomaly extraction unit, and a path prediction unit;

[0103] The alignment and association unit is used to align and associate knowledge graphs with high similarity in real time by taking advantage of the exponential growth of the spatial capacity of hyperbolic space as the radius increases, and by using graph neural networks to capture the high-order proximity of entities through the neighborhood information of entities.

[0104] The anomaly extraction unit is used to automatically extract anomaly knowledge graphs that meet the threshold stratification when knowledge graph differences occur;

[0105] The path prediction unit is used to calculate the deviation angle of the anomaly knowledge graph in real time. The calculation formula is as follows:

[0106]

[0107] Where α is the angle between the path distance of the abnormal knowledge graph and the positive direction of the X-axis, θ is the angle between the movement distance of the abnormal graph and the positive direction of the Y-axis, A is the height of the abnormal knowledge graph, Δx is the vector of the abnormal knowledge graph in the X-axis direction, Δy is the vector of the abnormal knowledge graph in the Y-axis direction, and vector = 0 indicates that the abnormal knowledge graph path is normal, while vector ≠ 0 indicates that the abnormal knowledge graph path deviates.

[0108] By performing semantic disambiguation on entities in real time, knowledge fusion is achieved and automatically labeled and sorted. Pattern alignment is performed on entities that cannot complement each other, enabling timely semantic disambiguation. At the same time, through the mutual association between entities and the real-time prediction of abnormal entity paths, knowledge fusion is made easier to align and complement each other, and entity paths in the knowledge graph can be predicted in advance.

[0109] Example 3

[0110] Please refer to Figure 4 As shown: Based on Embodiment 1, the entity recognition module includes a two-layer learning module, a label prediction module, and an entity recognition module;

[0111] The two-layer learning module includes a bidirectional GRU neural network unit and a multi-source attention mechanism unit;

[0112] Bidirectional GRU neural network units are used to learn local information in a knowledge graph through a bidirectional GRU neural network.

[0113] The multi-source attention mechanism unit is used to learn global information based on local information through BIGRU LAN.

[0114] The label prediction module includes a local progressive unit and a label prediction unit;

[0115] Locally progressive units are used to learn progressive global information through a locally progressive mode.

[0116] The label prediction unit is used to predict labels based on the learned information.

[0117] The entity recognition module includes an output combination unit, an entity recognition unit, and a recognition verification unit;

[0118] The output combining unit is used to combine the hidden state and the output of the attention mechanism;

[0119] The entity recognition unit is used to recognize entities based on the label prediction results. The recognition method is as follows:

[0120] Step 1 (1): First, calculate the sum of the values ​​of the first layer and all candidate labels x. l =x1 l x2 l ,…,x|L| l An attention matrix α is generated through attention, which consists of the latent label distribution of each knowledge graph. Q = H is defined as follows. l K = V = x l ,in This uses a tag set representation, where |L| represents the total number of tags. Note that the output of matrix a is calculated as follows:

[0121] H l=attention(Q, K, V) = aV

[0122]

[0123] Step 1 (2): The label attention inference layer uses a multi-head attention mechanism. The calculation formula for multi-head attention is as follows:

[0124] H l = concat(head, ..., head) K )+H l

[0125] head i =attention(QW i Q KW i K VW i V )

[0126] Among them, W i Q W i K and W i V Here, k represents the parameters that need to be learned during the training process, and k is the number of threads.

[0127] Step 1 (3): The final output of the BIGRU LAN layer combines the hidden state of BIGRU and the output of the attention mechanism. Let H represent the final output, i.e., H = [H 1 H l ];

[0128] Step IV: The output of the first layer is then used as the input to the second BIGRULAN layer. The second BIGRULAN layer then predicts the labels of the knowledge graph based on the attention weights. The prediction formula is as follows:

[0129]

[0130] yi = argmax j (yi 1 , ..., yi n )

[0131] Among them, yi n Let represent the nth candidate label of the i-th knowledge graph, and yi represent the predicted label of the i-th knowledge graph in the sequence;

[0132] The recognition and verification unit is used to verify the entity recognition results in real time. It compares the recognition results with the knowledge graph set by the system. If there is a difference, it indicates that the recognition result is invalid. The system will issue a voice reminder and repeat the entity recognition process. If there is no difference, it indicates that the recognition result is valid.

[0133] By using a bidirectional GRU neural network layer and a multi-head attention mechanism layer, local information is first learned, and then global information is learned based on the local information, thus achieving distributed learning, reducing the scale difference between multi-source knowledge graphs, and performing label prediction based on global information. The combination of hidden state and attention mechanism output enables efficient entity recognition, and can optimize decision analysis to improve the path search efficiency based on the full-element time-varying knowledge graph.

[0134] This invention discloses a decision analysis method based on a full-element time-varying knowledge graph. First, it enters a hyperbolic capture end, where data from different sources is collected in real time, describing the same fact from different perspectives to form a heterogeneous knowledge graph. This graph is then used for real-time fact complementation and anomaly graph differentiation. By collecting data from different sources and describing the same fact from different perspectives, a heterogeneous knowledge graph is formed, and fact complementation is performed. This allows for fact complementation between different knowledge graphs during decision analysis, ensuring the preservation of complex hierarchical structures in a linear space and achieving automatic complementation. Furthermore, by identifying heterogeneous knowledge and anomaly graphs, anomalies in the knowledge graph decision analysis are extracted and highlighted in real time, improving the timeliness and accuracy of the knowledge graph. Next, it enters a knowledge fusion end, where semantic disambiguation of entities is performed in real time to achieve knowledge fusion and automatic identification and sorting. Pattern alignment is performed on entities that cannot be complemented, enabling inter-entity association. Anomaly entities are extracted, and their paths are predicted in real time. This process of semantic disambiguation of entities ensures real-time knowledge fusion and automatic identification and sorting. The system integrates and automatically identifies and sorts knowledge, aligns patterns for entities that cannot complement each other, and performs timely semantic disambiguation. It also predicts abnormal entity paths in real time by identifying inter-entity relationships, facilitating knowledge integration and complementarity, and enabling advance prediction of entity paths in the knowledge graph. At the entity recognition end, a bidirectional GRU neural network layer and a multi-head attention mechanism layer are used to first learn local information, then learn global information from that local information, achieving distributed learning. Label prediction is then performed based on the global information, combining hidden states and attention mechanism outputs to achieve efficient entity recognition. This process, along with optimizing decision analysis, improves path search efficiency based on a full-element time-varying knowledge graph.

[0135] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of this invention is defined by the appended claims and their equivalents.

Claims

1. A decision analysis method based on a full-factor time-varying knowledge graph, characterized in that, The method includes the following implementation steps: Step 1: Enter the hyperbolic capture terminal, collect data from different sources in real time, describe the same fact from different perspectives, form a heterogeneous knowledge graph, complement the facts in real time, and distinguish abnormal graphs; Step 2: Enter the knowledge fusion terminal, perform semantic disambiguation on entities in real time, realize knowledge fusion and automatic identification and sorting, perform pattern alignment on entities that cannot complement each other, realize the mutual association between entities, extract abnormal entities, and predict the path of abnormal entities in real time. Step 3: Enter the entity recognition end. Through the bidirectional GRU neural network layer and the multi-head attention mechanism layer, local information is first learned, and then global information is learned based on the local information to achieve distributed learning. Label prediction is performed based on the global information. The hidden state and the output of the attention mechanism are combined to achieve efficient entity recognition.

2. The method according to claim 1, characterized in that: The hyperbolic capture end includes a hyperbolic receiving module, a hyperbolic complementary module, and a heterogeneous differentiation module; The hyperbolic receiver module includes a multi-dimensional acquisition unit and a multi-dimensional description unit; The multi-dimensional acquisition unit is used to acquire knowledge graphs from different angles in real time using a data acquisition instrument; The multidimensional description unit is used to track and describe the same knowledge fact from different perspectives in real time using a data tracker.

3. The method according to claim 2, characterized in that: The hyperbolic complementary module includes a hyperbolic construction unit and a knowledge complementarity unit; The hyperbolic construction unit is used to construct a decision analysis model for time-varying knowledge graphs using a hyperbolic neural network; The knowledge complementation unit is used to complement the differences in the knowledge graph of the decision analysis model, as follows: Step 1: Calculate the affinity matrix in the decision analysis model at the current time step to measure the similarity of node attribute information. The formula is as follows: Where C represents the affinity matrix in the decision analysis model at the current moment, which measures the similarity of node attribute information, and x i and y i This represents the i-th vector in the knowledge graph information and the knowledge graph information in the decision analysis model at the current moment; Step 2: When two vectors are more similar, the closer their similarity value is to 1, the closer the knowledge graphs are; conversely, the closer they are, the greater the differences between the knowledge graphs.

4. The method according to claim 3, characterized in that: The heterogeneous differentiation module includes a heterogeneous knowledge unit, an anomaly differentiation unit, and an anomaly alert unit; The heterogeneous knowledge unit is used to identify one-quarter of the knowledge graphs with a similarity value less than or equal to 1 as abnormal knowledge. The anomaly differentiation unit is used to perform threshold stratification based on the similarity of anomaly knowledge; The anomaly alert unit is used to report the hierarchical knowledge graph of the anomaly knowledge to the system and issue a voice alarm.

5. The method according to claim 1, characterized in that: The knowledge fusion terminal includes a semantic disambiguation module, a fusion ranking module, and a path prediction module; The semantic disambiguation module is used to measure the degree of association between semantics and context in the knowledge graph by calculating the mutual information between them in real time, as follows: Step 1: For each word meaning, calculate its mutual information with each word in the context; Step II: Accumulate all mutual information values ​​to obtain the total score, and sort the semantics of the current time step by number according to the sequence number; Step 3: Select the word meaning with the highest total score at the current moment as the final disambiguation result, and then proceed to the next semantic disambiguation step simultaneously.

6. The method according to claim 5, characterized in that: The fusion sorting module includes a knowledge fusion unit and an identifier sorting unit; The knowledge fusion unit is used to achieve real-time fusion of the knowledge graph by fusing the knowledge graph at the current moment, and to record it in real time by a data recorder. The identification and sorting unit is used to color-code the knowledge graph after knowledge fusion and to arrange it by sequence number.

7. The method according to claim 6, characterized in that: The path prediction module includes an alignment and association unit, an anomaly extraction unit, and a path prediction unit. The alignment and association unit is used to align and associate knowledge graphs with high similarity in real time by utilizing the fact that the spatial capacity of hyperbolic space increases exponentially with the radius, and by using graph neural networks to capture the high-order proximity of entities through the neighborhood information of entities. The anomaly extraction unit is used to automatically extract anomaly knowledge graphs that meet the threshold stratification when knowledge graph differences occur. The path prediction unit is used to calculate the deviation angle of the abnormal knowledge graph in real time. The calculation formula is as follows: Where α is the angle between the path distance of the abnormal knowledge graph and the positive direction of the X-axis, θ is the angle between the movement distance of the abnormal graph and the positive direction of the Y-axis, A is the height of the abnormal knowledge graph, Δx is the vector of the abnormal knowledge graph in the X-axis direction, Δy is the vector of the abnormal knowledge graph in the Y-axis direction, and vector = 0 indicates that the abnormal knowledge graph path is normal, while vector ≠ 0 indicates that the abnormal knowledge graph path deviates.

8. The method according to claim 1, characterized in that: The entity recognition module includes a two-layer learning module, a label prediction module, and an entity recognition module. The two-layer learning module includes a bidirectional GRU neural network unit and a multi-source attention mechanism unit; The bidirectional GRU neural network unit is used to learn local information in the knowledge graph through the bidirectional GRU neural network. The multi-source attention mechanism unit is used to learn global information based on local information through BIGRU LAN.

9. The method according to claim 8, characterized in that: The label prediction module includes a local progressive unit and a label prediction unit; The local progressive unit is used to learn progressive global information through a local progressive mode; The label prediction unit is used to predict labels based on the learned information.

10. The method according to claim 9, characterized in that: The entity recognition module includes an output combining unit, an entity recognition unit, and a recognition verification unit; The output combining unit is used to combine the outputs of the hidden state and the attention mechanism. The entity recognition unit is used to perform entity recognition based on the label prediction results. The recognition method is as follows: Step 1 (1): First, calculate the sum of the values ​​of the first layer and all candidate labels x. l =x1 l x2 l ,…,x|L| l An attention matrix α is generated through attention, which consists of the latent label distribution of each knowledge graph. Q = H is defined as follows. l K = V = x l ,in This uses a tag set representation, where |L| represents the total number of tags. Note that the output of matrix a is calculated as follows: H l =attention(Q,K,V)=aV Step 1 (2): The label attention inference layer uses a multi-head attention mechanism. The calculation formula for multi-head attention is as follows: H l =concat(head,…,head K )+H l head i =attention(QW i Q ,KW i K ,VW i V ) Among them, W i Q W i K and W i V Here, k represents the parameters that need to be learned during the training process, and k is the number of threads. Step 1 (3): The final output of the BIGRU LAN layer combines the hidden state of BIGRU and the output of the attention mechanism. Let H represent the final output, i.e., H = [H 1 H l ]; Step IV: The output of the first layer is then used as the input to the second BIGRU LAN layer. The second BIGRU LAN layer then predicts the labels of the knowledge graph based on the attention weights. The prediction formula is as follows: yi = argmax j (yi 1 , ..., yi n ) Among them, yi n Let represent the nth candidate label of the i-th knowledge graph, and yi represent the predicted label of the i-th knowledge graph in the sequence; The identification verification unit is used to verify the entity recognition result in real time. It compares the recognition result with the knowledge graph set by the system. If there is a difference in the comparison, it indicates that the recognition result is invalid. The system is then notified to issue a voice reminder and repeat the entity recognition process. If there is no difference in the comparison, it indicates that the recognition result is valid.

Images