A method for multi-department government affair knowledge graph construction and agent collaborative decision-making based on asynchronous federated learning
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- CHINA UNICOM XIONGAN IND INTERNET CO LTD
- Filing Date
- 2026-04-09
- Publication Date
- 2026-07-03
AI Technical Summary
Existing technologies for sharing government data suffer from privacy risks, difficulties in entity mapping, blurred decision boundaries, and insufficient continuous iteration capabilities. In particular, asynchronous federated learning schemes struggle to achieve cross-departmental knowledge fusion and agent decision-making flexibility.
A method for constructing a multi-departmental government knowledge graph based on asynchronous federated learning is adopted. By deploying CLIP models in each government department for data processing, cross-departmental entity matching is performed using graph neural networks and seedless entity alignment technology. Combined with intelligent agents, two-factor decision boundary judgment is performed, and automated process generation is achieved through visualization tools and RPA rule base.
It achieves semantic-level fusion of cross-departmental data, solves the problems of privacy leakage risks and entity mapping difficulties, improves the rationality of human-machine collaboration and the degree of process automation, and adapts to the changes in complex government processes.
Smart Images
Figure CN122334445A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of privacy computing technology, and in particular to a method for constructing a multi-departmental government knowledge graph and making collaborative decisions with intelligent agents based on asynchronous federated learning. Background Technology
[0002] With the deepening of digital government construction, a large amount of heterogeneous, multi-source, and multi-modal data resources have been accumulated in the field of government affairs.
[0003] However, this data is often scattered across different departments and systems, creating significant information silos. Traditional centralized data sharing models face the following challenges: 1) Data security and privacy protection: Government data contains a large amount of sensitive information, and direct sharing can easily lead to privacy leaks. 2) Data heterogeneity: Data from different departments differs significantly in format, structure, and semantics. For example, the naming rules for license plate data in the public security system may differ from those in the traffic management department, leading to difficulties in entity mapping. 3) Ambiguous decision-making boundaries: The role of AI in government decision-making is still unclear, making it difficult to define the applicable scenarios for AI suggestions, manual review, and automated execution. 4) Insufficient continuous iteration capabilities: Policies, regulations, and business processes are constantly changing, and existing systems lack efficient cross-departmental collaborative update mechanisms. Federated learning, due to its characteristic of keeping the data stationary while allowing the model to move, has become an important direction for solving these problems. It allows each participant to train the model locally, exchanging only model parameters or intermediate representations, thereby avoiding the leakage of raw data. However, existing federated learning solutions face the following limitations in government scenarios: the synchronous update mechanism is difficult to adapt to large-scale distributed deployment; there is a lack of effective seedless entity alignment technology, making it impossible to achieve cross-departmental knowledge fusion while protecting privacy; the agent decision boundary adjustment mechanism is not flexible enough and is difficult to adapt to complex government processes; and there is a lack of closed-loop linkage between knowledge graph updates and automated processes. Summary of the Invention
[0004] To overcome the shortcomings of existing technologies, the purpose of this invention is to provide a method for constructing a multi-departmental government knowledge graph and making collaborative decisions with intelligent agents based on asynchronous federated learning, thereby solving the problems of privacy leakage risks, difficulties in entity mapping, and ambiguous decision boundaries that existing methods are prone to.
[0005] To achieve the above objectives, the present invention provides the following solution: A method for constructing a multi-sectoral government knowledge graph and enabling collaborative decision-making by intelligent agents based on asynchronous federated learning includes: Deploy a corresponding CLIP model at each government department business node, and use the CLIP model to process the modal data of the target node to obtain CLIP joint embedding features; Asynchronous federated learning is used to perform cross-departmental collaborative updates of the CLIP model corresponding to all nodes; By using graph neural network technology and seedless entity alignment technology, relational reasoning and cross-departmental entity matching are performed on the CLIP joint embedding features to obtain a multimodal knowledge graph; Based on the multimodal knowledge graph, an agent fine-tuned with domain knowledge performs a two-factor decision boundary judgment on the CLIP joint embedding features corresponding to the current task data, obtains a comprehensive judgment result, and matches the decision execution path corresponding to the comprehensive judgment result; the decision execution path includes one of: automatic execution, auxiliary suggestion, and mandatory manual review; The multimodal knowledge graphs corresponding to each department are merged into a domain-level large graph. The potential conflict points of the domain-level large graph are highlighted using visualization tools, and the feedback on the potential conflict points is updated to the exception condition library. The domain-level large map and the business rules that do not conflict or require manual approval in the incremental data within the set window of the exception condition library are injected into the RPA rule library, and the corresponding automated process is generated based on the RPA rule library; the set window is not less than one month.
[0006] Preferably, a corresponding CLIP model is deployed at each government department business node, and the CLIP model is used to process the modal data of the target node to obtain CLIP joint embedding features, including: Receive raw multimodal data; the raw multimodal data includes: text, images, and time-series data; The original multimodal data is standardized and preprocessed, and the time-series data is transformed into a visual image representation; The CLIP model is used to extract features from the preprocessed original multimodal data in the same semantic space of each department, thereby obtaining the CLIP joint embedding features.
[0007] Preferably, asynchronous federated learning is used to perform cross-departmental collaborative updates of the CLIP model corresponding to all nodes, including: The CLIP model for each department node is trained independently to obtain local model updates; The local model is updated and uploaded to the coordination center; The local model update is aggregated according to the dynamic weight algorithm to obtain the updated global model, and the global model is distributed to each department node.
[0008] Preferably, the expression for the dynamic weighting algorithm is: ;in, For the first Aggregate weights of each department; To adjust hyperparameters; Rate the data quality; It is a countdown to the time between the last update and the current update.
[0009] Preferably, the data transmission process of the coordination center includes: Extract the knowledge patterns and reasoning rules of the CLIP model from each government node to obtain structured knowledge fragments, and upload the structured knowledge fragments to the coordination center; The structured knowledge fragments uploaded by each government node are horizontally merged to obtain a unified knowledge representation; The unified knowledge representation is synchronously transmitted to each government node using the LoRa module.
[0010] Preferably, graph neural network technology and seedless entity alignment technology are used to perform relational reasoning and cross-departmental entity matching on the CLIP joint embedding features to obtain a multimodal knowledge graph, including: Construct an initial entity representation based on the CLIP joint embedding features; The initial representation of the entity is fused using a cross-modal attention mechanism to obtain multimodal features; Based on the multimodal features, a knowledge graph is constructed using graph neural network technology to obtain a departmental knowledge graph; the departmental knowledge graph adopts a triplet format. Seedless entity alignment technology is used to perform entity matching on the departmental knowledge graphs of each department to obtain the multimodal knowledge graph with cross-departmental matching.
[0011] Preferably, the expression for the entity alignment score of the knowledge graphs of different departments is: ;in, Assign an alignment score to the entity; , These are the entities of the first and second maps to be aligned, respectively. for The corresponding embedding vector; for The corresponding embedding vector.
[0012] Preferably, based on the multimodal knowledge graph, an agent fine-tuned with domain knowledge performs a two-factor decision boundary judgment on the CLIP joint embedding features corresponding to the current task data, obtains a comprehensive judgment result, and matches the decision execution path corresponding to the comprehensive judgment result, including: The agent was selected after being fine-tuned with domain knowledge; the selection criterion for the agent was its performance score on a preset test set, ranking in the top 10%. Set judgment rules; the judgment rules include: marking the model confidence level as requiring manual intervention when it is lower than a set threshold, and triggering manual review when the knowledge graph rule matching degree is lower than a preset threshold; According to the determination rule, the CLIP joint embedding feature corresponding to the current task data is input into the agent to determine the execution path, and the determined execution path is obtained.
[0013] Preferably, the model confidence score is measured using the maximum class probability distribution value or prediction entropy to measure uncertainty; the knowledge graph rule matching degree is measured using Jaccard similarity or cosine similarity.
[0014] Preferably, it further includes: Set the full federation aggregation cycle; When the full federated aggregation cycle is reached or business adjustments are made, the GraphRAG engine is used to extract new input entity relationships to obtain incremental relationships; The multimodal knowledge graph is updated using the incremental relationships processed with differential privacy.
[0015] The present invention discloses the following technical effects: This invention provides a method for constructing a multi-departmental government knowledge graph and making collaborative decisions with intelligent agents based on asynchronous federated learning. By using seedless entity alignment and multimodal embedding technology, it solves the problems of privacy leakage risks and entity mapping difficulties in existing methods, and realizes semantic-level fusion of heterogeneous data. Through a two-factor threshold determination mechanism, it solves the problem of ambiguous decision boundaries in existing methods and improves the rationality of human-machine collaboration. Attached Figure Description
[0016] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0017] Figure 1 A schematic diagram illustrating the process of constructing a multi-departmental government knowledge graph and making collaborative decisions with intelligent agents based on asynchronous federated learning, provided for an embodiment of the present invention; Figure 2 The human-machine collaborative optimization closed-loop diagram provided in the embodiments of the present invention. Detailed Implementation
[0018] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0019] The purpose of this invention is to provide a method for constructing a multi-departmental government knowledge graph and making collaborative decisions with intelligent agents based on asynchronous federated learning, which solves the problems of privacy leakage risks, difficulty in entity mapping, and ambiguous decision boundaries in existing methods.
[0020] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.
[0021] Figure 1 This is a schematic diagram illustrating the process of constructing a multi-departmental government knowledge graph and making collaborative decisions with intelligent agents based on asynchronous federated learning, as provided in an embodiment of the present invention. Figure 2 The human-machine collaborative optimization closed-loop diagram provided in the embodiments of the present invention is as follows: Figure 1 and Figure 2 As shown, this invention provides a method for constructing a multi-departmental government knowledge graph and enabling collaborative decision-making by intelligent agents based on asynchronous federated learning, including: Step 100: Deploy the corresponding CLIP model at each government department business node, and use the CLIP model to process the modal data of the target node to obtain CLIP joint embedding features; Step 200: Use asynchronous federated learning to perform cross-departmental collaborative updates on the CLIP model corresponding to all nodes; Step 300: Using graph neural network technology and seedless entity alignment technology, perform relational reasoning and cross-departmental entity matching on the CLIP joint embedding features to obtain a multimodal knowledge graph; Step 400: Based on the multimodal knowledge graph, the agent, fine-tuned by domain knowledge, performs a two-factor decision boundary judgment on the CLIP joint embedding features corresponding to the current task data to obtain a comprehensive judgment result, and matches the decision execution path corresponding to the comprehensive judgment result; the decision execution path includes one of: automatic execution, auxiliary suggestion, and forced manual review; Step 500: Merge the multimodal knowledge graphs corresponding to each department into a domain-level large graph, use visualization tools to highlight the potential conflict points of the domain-level large graph, and update the feedback on the potential conflict points to the exception condition library; Step 600: Inject the business rules that do not conflict or require manual approval in the incremental data within the set window of the domain-level large map and the exception condition library into the RPA rule library, and generate the corresponding automated process for automatic execution based on the RPA rule library; the set window is not less than one month.
[0022] Furthermore, a corresponding CLIP model is deployed at each government department's business node. The CLIP model is then used to process the modal data of the target node to obtain CLIP joint embedding features, including: Receive raw multimodal data; the raw multimodal data includes: text, images, and time-series data; The original multimodal data is standardized and preprocessed, and the time-series data is transformed into a visual image representation; The CLIP model is used to extract features from the preprocessed original multimodal data in the same semantic space of each department, thereby obtaining the CLIP joint embedding features.
[0023] Furthermore, asynchronous federated learning is used to perform cross-departmental collaborative updates of the CLIP model corresponding to all nodes, including: The CLIP model for each department node is trained independently to obtain local model updates; The local model is updated and uploaded to the coordination center; The local model update is aggregated according to the dynamic weight algorithm to obtain the updated global model, and the global model is distributed to each department node.
[0024] Specifically, the expression for the dynamic weighting algorithm is: ;in, For the first Aggregate weights of each department; To adjust hyperparameters; Rate the data quality; It is a countdown to the time between the last update and the current update.
[0025] Furthermore, the data transmission process of the coordination center includes: Extract the knowledge patterns and reasoning rules of the CLIP model from each government node to obtain structured knowledge fragments, and upload the structured knowledge fragments to the coordination center; The structured knowledge fragments uploaded by each government node are horizontally merged to obtain a unified knowledge representation; The unified knowledge representation is synchronously transmitted to each government node using the LoRa module.
[0026] Preferably, graph neural network technology and seedless entity alignment technology are used to perform relational reasoning and cross-departmental entity matching on the CLIP joint embedding features to obtain a multimodal knowledge graph, including: Construct an initial entity representation based on the CLIP joint embedding features; The initial representation of the entity is fused using a cross-modal attention mechanism to obtain multimodal features; Based on the multimodal features, a knowledge graph is constructed using graph neural network technology to obtain a departmental knowledge graph; the departmental knowledge graph adopts a triplet format. Seedless entity alignment technology is used to perform entity matching on the departmental knowledge graphs of each department to obtain the multimodal knowledge graph with cross-departmental matching.
[0027] Specifically, the expression for the entity alignment score of the knowledge graphs of different departments is as follows: ;in, Assign an alignment score to the entity; , These are the entities of the first and second maps to be aligned, respectively. for The corresponding embedding vector; for The corresponding embedding vector.
[0028] Furthermore, based on the multimodal knowledge graph, the agent, fine-tuned with domain knowledge, performs a two-factor decision boundary judgment on the CLIP joint embedding features corresponding to the current task data, obtains a comprehensive judgment result, and matches the decision execution path corresponding to the comprehensive judgment result, including: The agent was selected after being fine-tuned with domain knowledge; the selection criterion for the agent was its performance score on a preset test set, ranking in the top 10%. Set judgment rules; the judgment rules include: marking the model confidence level as requiring manual intervention when it is lower than a set threshold, and triggering manual review when the knowledge graph rule matching degree is lower than a preset threshold; According to the determination rule, the CLIP joint embedding feature corresponding to the current task data is input into the agent to determine the execution path, and the determined execution path is obtained.
[0029] Optionally, the model confidence score uses the maximum class probability distribution value or prediction entropy to measure uncertainty; the knowledge graph rule matching degree uses Jaccard similarity or cosine similarity.
[0030] Furthermore, it also includes: Set the full federation aggregation cycle; When the full federated aggregation cycle is reached or business adjustments are made, the GraphRAG engine is used to extract new input entity relationships to obtain incremental relationships; The multimodal knowledge graph is updated using the incremental relationships processed with differential privacy.
[0031] Preferably, this embodiment provides a method for constructing a multi-departmental government knowledge graph and enabling collaborative decision-making by intelligent agents based on artificial intelligence. This method extracts multimodal features through local edge nodes and uses an asynchronous federated learning framework to aggregate cross-departmental model parameters. It generates joint embedding vectors in a unified semantic space using the CLIP model and constructs departmental-level knowledge subgraphs using graph neural network technology. It introduces zero-prior cross-department entity matching technology to complete cross-departmental entity matching without the need for pre-labeled mapping tables. It deploys intelligent agents based on large language models to perform business logic analysis and decision simulation, dynamically determining AI execution permissions. A closed-loop manual verification and feedback mechanism is established, injecting stable rules into the RPA automated process while updating the exception condition library to store uncertain cases, forming a continuously optimized knowledge governance system.
[0032] Specifically, local data processing and feature extraction are performed. Optionally, local data processing endpoints can be deployed at each government department's business node to perform the following operations: receive raw multimodal data, including text, images, and time-series data; perform standardized preprocessing, transforming time-series data into visual image representations (such as spectrograms and trend charts); input all modal data into the government-optimized CLIP model, extracting joint embedding feature vectors under a unified semantic space; and apply feature compression techniques to effectively compress data dimensions, retaining valid information for subsequent modeling.
[0033] Image encoder: Text encoder: Similarity score: Training objective (contrastive loss): ;in, Input image; For the corresponding text description; These are image and text encoders, respectively. These are the embedding vectors for the image and the text, respectively; A similarity score; This is the temperature coefficient.
[0034] .
[0035] in, (e.g., PCA) or non-linear coding: (e.g., self-encoder) This is the original high-dimensional feature vector; This is the compressed low-dimensional representation; Compression function; Projection matrix; It is a parameterized encoding network.
[0036] Furthermore, asynchronous federated learning model updates are implemented. An asynchronous federated learning process is executed to achieve cross-departmental collaborative modeling: each department node independently trains its local model and generates local model updates; these local model updates are uploaded to the coordination center, with the upload timing determined by the readiness of the local data, requiring no global synchronization; the coordination center receives update requests from any node and immediately performs aggregation operations; during aggregation, a dynamic weighting algorithm is used, comprehensively considering departmental data quality and update timeliness, assigning different contribution weights; the updated global model is distributed back to participating nodes for the next round of local training.
[0037] .
[0038] in, For the first Aggregate weights of each department; Rate the data quality of this department (e.g., completeness, accuracy); The highest data quality among all departments; It is the countdown to the last update (i.e.) This reflects the timeliness of the update; This is the maximum allowed interval time; Adjusting hyperparameters to control the relative importance of quality and timeliness.
[0039] Preferably, a multimodal knowledge graph is constructed. Departmental-level knowledge subgraphs are built based on the local model output: Initial entity representations are constructed based on the CLIP joint embedding features obtained in the preceding steps, and relation reasoning and subgraph generation are performed using graph neural network technology; a cross-modal attention mechanism is designed to achieve deep fusion of multimodal features; an internal departmental knowledge graph is constructed, containing entity, attribute, and relation triples; a seedless entity alignment technique (such as TransAlign) is used to achieve cross-departmental entity matching without pre-labeled alignment samples; knowledge credibility is quantified based on a Bayesian network to support uncertain reasoning. For entities from two knowledge graphs… Its alignment score is defined as: .
[0040] in, , These are the embeddings in their respective graphs. The training objective is to minimize the positive example distance: ; For alignment scores, the closer to 0, the more likely alignment is. For a known set of aligned entity pairs (even without a seed, a small amount of weak supervision can still be used); For encoders that are shared or mapped across graphs.
[0041] .
[0042] Alternatively, a Bayesian network joint distribution decomposition can be used: .
[0043] Credibility is defined as: ;in, This is a hypothetical fact (such as a knowledge triple being true). For observational evidence; Here, is the posterior probability, representing the probability that the hypothesis is true given the evidence. For nodes The set of parent nodes; For rules Credibility.
[0044] .
[0045] If we assume the image features are... Text features are Then cross-modal attention can be expressed as: .
[0046] in, These are the query, key, and value matrices, respectively. These are the learnable linear transformation matrices; The dimension of the key vector; Attention output for image-to-text conversion enables cross-modal information fusion. Specifically, the process involves intelligent agent collaborative decision-making and boundary determination. Intelligent agents are deployed for business logic analysis and decision simulation: the large language model, based on feature vectors extracted by CLIP and their contextual metadata, understands the semantics of the business process and identifies potential patterns and exceptions. Preferably, the large language model is a Transformer architecture model fine-tuned with domain knowledge, selected based on the top 10% of the intelligent agent call performance scores on a preset test set, to ensure optimal responsiveness and logical consistency in specific business scenarios. Each intelligent agent integrates a dialogue management module, the core of the large language model, and tool interfaces; it inputs current task data and executes an AI-simulated decision-making process; it calculates the two-factor decision boundary: analyzes the model confidence level, and marks it as requiring manual intervention when it falls below a set threshold; it compares the consistency between local rules and the global knowledge graph, and triggers manual review when the knowledge graph rule matching degree falls below a preset threshold; and it determines the execution path based on the comprehensive judgment result: automatic execution, assisted suggestions, or mandatory manual review.
[0047] Furthermore, let the local rule set be... The global rule set is The matching degree can be defined as Jaccard similarity: .
[0048] Alternatively, cosine similarity can be used (if the rules are embedded as vectors): .
[0049] in, These are the local and global sets of knowledge rules (such as sets of triples); The value represents the degree of consistency of the rules; a higher value indicates greater consistency. It is a vector representation of the rule set (e.g., obtained through average pooling).
[0050] Let the probability distribution of the model output class be... The confidence level is defined as: Alternatively, predictive entropy can be used to measure uncertainty: The corresponding confidence level is: ;in, This is the probability vector output by the softmax function; Total number of categories; This is the maximum class probability, often used for confidence estimation; For predicting entropy, a larger value indicates greater uncertainty. Furthermore, human-machine collaboration optimization and process integration are implemented. A feedback loop is established to continuously optimize system performance: departmental sub-maps are merged into a domain-level large map, and potential conflict points are highlighted using visualization tools; a manual verification interface is provided, supporting expert review and correction of map consistency; human feedback is collected, encrypted, and used to update the exception condition library; verified and stable business rules are injected into the RPA rule base via API to generate executable automated processes; multiple triggering methods are supported, including timed triggering, event triggering, and API calls, to achieve automated process execution.
[0051] Preferably, continuous iteration and version management are implemented. This maintains the system's long-term evolution capability: regular full-scale federated aggregation cycles are performed, and the contribution ratios of each department are adjusted based on dynamic weights; when significant policy changes or business adjustments occur, an event-driven incremental update mechanism is initiated; newly released entity relationships are extracted using the GraphRAG engine, and the knowledge graph is partially updated after differential privacy processing; a Git-like version control system is used to record the changes in each update, supporting the rollback of graph and process versions.
[0052] Optionally, a data transmission mechanism is employed. This embodiment adopts a data transmission paradigm based on local model knowledge and rules. After each government node completes model training locally, it does not directly transmit gradients or parameters, but instead extracts the knowledge patterns and reasoning rules contained in the local model to form structured knowledge fragments. After receiving the knowledge fragments from each node, the coordination center performs a horizontal merging operation to integrate them into a unified knowledge representation. Preferably, these knowledge fragments can be transmitted over low power and wide area via a LoRa module, which is particularly suitable for horizontal knowledge merging in edge-side resource-constrained scenarios. The LoRa module serves as an optional communication carrier here, enabling long-distance, low-power knowledge synchronization. Furthermore, zero-knowledge proofs can be attached to the knowledge fragments before transmission to verify the legitimacy of their source.
[0053] The beneficial effects of this invention are as follows: This invention achieves semantic-level fusion of heterogeneous data through seedless entity alignment and multimodal embedding technology, breaking down data silos; improves the rationality of human-machine collaboration through a two-factor threshold determination mechanism; and enhances the automation level of processes by automatically mapping stable business rules into RPA action sequences.
[0054] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. The same or similar parts between the various embodiments can be referred to each other.
[0055] This document uses specific examples to illustrate the principles and implementation methods of the present invention. The descriptions of the above embodiments are only for the purpose of helping to understand the method and core ideas of the present invention. Furthermore, those skilled in the art will recognize that, based on the ideas of the present invention, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of the present invention.
Claims
1. A method for multi-department government knowledge graph construction and agent collaborative decision-making based on asynchronous federated learning, characterized in that, include: Deploy a corresponding CLIP model at each government department business node, and use the CLIP model to process the modal data of the target node to obtain CLIP joint embedding features; Asynchronous federated learning is used to perform cross-departmental collaborative updates of the CLIP model corresponding to all nodes; By using graph neural network technology and seedless entity alignment technology, relational reasoning and cross-departmental entity matching are performed on the CLIP joint embedding features to obtain a multimodal knowledge graph; Based on the multimodal knowledge graph, the agent, fine-tuned with domain knowledge, performs a two-factor decision boundary judgment on the CLIP joint embedding features corresponding to the current task data, obtains a comprehensive judgment result, and matches the decision execution path corresponding to the comprehensive judgment result. The decision-making execution path includes one of the following: automatic execution, auxiliary suggestions, or mandatory manual review; The multimodal knowledge graphs corresponding to each department are merged into a domain-level large graph. The potential conflict points of the domain-level large graph are highlighted using visualization tools, and the feedback on the potential conflict points is updated to the exception condition library. The domain-level large map and the business rules that do not conflict or require manual approval in the incremental data within the set window of the exception condition library are injected into the RPA rule library, and the corresponding automated process is generated based on the RPA rule library; the set window is not less than one month.
2. The method for constructing a multi-departmental government knowledge graph and making collaborative decisions with intelligent agents based on asynchronous federated learning according to claim 1, characterized in that, A corresponding CLIP model is deployed at each government department's business node. The CLIP model is then used to process the modal data of the target node to obtain CLIP joint embedding features, including: Receive raw multimodal data; the raw multimodal data includes: text, images, and time-series data; The original multimodal data is standardized and preprocessed, and the time-series data is transformed into a visual image representation; The CLIP model is used to extract features from the preprocessed original multimodal data in the same semantic space of each department, thereby obtaining the CLIP joint embedding features.
3. The method for constructing a multi-departmental government knowledge graph and making collaborative decisions with intelligent agents based on asynchronous federated learning according to claim 1, characterized in that, Asynchronous federated learning is used to perform cross-departmental collaborative updates of the CLIP model corresponding to all nodes, including: The CLIP model for each department node is trained independently to obtain local model updates; The local model is updated and uploaded to the coordination center; The local model update is aggregated according to the dynamic weight algorithm to obtain the updated global model, and the global model is distributed to each department node.
4. The method for constructing a multi-departmental government knowledge graph and making collaborative decisions with intelligent agents based on asynchronous federated learning according to claim 3, characterized in that, The expression for the dynamic weighting algorithm is: ;in, For the first Aggregate weights of each department; To adjust hyperparameters; Rate the data quality; It is a countdown to the time between the last update and the current update.
5. The method for constructing a multi-departmental government knowledge graph and making collaborative decisions with intelligent agents based on asynchronous federated learning according to claim 3, characterized in that, The data transmission process of the coordination center includes: Extract the knowledge patterns and reasoning rules of the CLIP model from each government node to obtain structured knowledge fragments, and upload the structured knowledge fragments to the coordination center; The structured knowledge fragments uploaded by each government node are horizontally merged to obtain a unified knowledge representation; The unified knowledge representation is synchronously transmitted to each government node using the LoRa module.
6. The method for constructing a multi-departmental government knowledge graph and making collaborative decisions with intelligent agents based on asynchronous federated learning according to claim 1, characterized in that, Using graph neural network technology and seedless entity alignment technology, relational reasoning and cross-departmental entity matching are performed on the CLIP joint embedding features to obtain a multimodal knowledge graph, including: Construct an initial entity representation based on the CLIP joint embedding features; The initial representation of the entity is fused using a cross-modal attention mechanism to obtain multimodal features; Based on the multimodal features, a knowledge graph is constructed using graph neural network technology to obtain a departmental knowledge graph; the departmental knowledge graph adopts a triplet format. Seedless entity alignment technology is used to perform entity matching on the departmental knowledge graphs of each department to obtain the multimodal knowledge graph with cross-departmental matching.
7. The method for constructing a multi-departmental government knowledge graph and making collaborative decisions with intelligent agents based on asynchronous federated learning according to claim 6, characterized in that, The expression for the entity alignment score of the knowledge graphs of different departments is as follows: ;in, Assign an alignment score to the entity; , These are the entities of the first and second maps to be aligned, respectively. for The corresponding embedding vector; for The corresponding embedding vector.
8. The method for constructing a multi-departmental government knowledge graph and making collaborative decisions with intelligent agents based on asynchronous federated learning according to claim 1, characterized in that, Based on the multimodal knowledge graph, an agent fine-tuned with domain knowledge performs a two-factor decision boundary judgment on the CLIP joint embedding features corresponding to the current task data, obtains a comprehensive judgment result, and matches the decision execution path corresponding to the comprehensive judgment result, including: The agent was selected after being fine-tuned with domain knowledge; the selection criterion for the agent was its performance score on a preset test set, ranking in the top 10%. Set judgment rules; the judgment rules include: marking the model confidence level as requiring manual intervention when it is lower than a set threshold, and triggering manual review when the knowledge graph rule matching degree is lower than a preset threshold; According to the determination rule, the CLIP joint embedding feature corresponding to the current task data is input into the agent to determine the execution path, and the determined execution path is obtained.
9. The method for constructing a multi-departmental government knowledge graph and making collaborative decisions with intelligent agents based on asynchronous federated learning according to claim 8, characterized in that, The model confidence level is measured by the maximum class probability distribution value or prediction entropy to measure uncertainty; the knowledge graph rule matching degree is measured by Jaccard similarity or cosine similarity.
10. The method for constructing a multi-departmental government knowledge graph and making collaborative decisions with intelligent agents based on asynchronous federated learning according to claim 1, characterized in that, Also includes: Set the full federation aggregation cycle; When the full federated aggregation cycle is reached or business adjustments are made, the GraphRAG engine is used to extract new input entity relationships to obtain incremental relationships; The multimodal knowledge graph is updated using the incremental relationships processed with differential privacy.