Dynamic knowledge graph construction system and method for multi-source data fusion

Abstract

This invention discloses a dynamic knowledge graph construction system and method based on multi-source data fusion, belonging to the field of computer technology. The system comprises a data layer, a knowledge layer, a technology layer, and an application layer. The data layer integrates multi-source heterogeneous data. The knowledge layer performs entity recognition, relation extraction, and knowledge fusion to generate a knowledge graph, and labels each piece of knowledge with a valid time attribute. The technology layer includes graph neural networks, a retrieval-enhanced generation framework, and an incremental learning engine. When the valid time endpoint of newly added knowledge is later than the latest valid time endpoint of existing knowledge, a new expert network is added, and the new data distribution is fitted through gradient updates. The application layer includes four intelligent agents: pre-sales, R&D, manufacturing, and service. This invention triggers graph neural network expansion through valid time attributes, achieving dynamic adaptation of the knowledge graph to the evolution of knowledge time. This overcomes the problems of traditional solutions struggling to manage knowledge timeliness and adapt models to time domain expansion, making it suitable for multi-source heterogeneous data processing scenarios such as manufacturing.

Description

Technical Field

[0001] This invention belongs to the field of computer big data and artificial intelligence technology, specifically involving the fusion of multi-source heterogeneous industrial data, the construction of time-series dynamic knowledge graphs, retrieval enhancement generation (RAG), incremental learning of graph neural networks (GNN), and collaborative decision-making technology of multi-business intelligent agents. It is applicable to the full-process knowledge asset accumulation and intelligent auxiliary decision-making of high-end manufacturing enterprises such as image modules. Background Technology

[0002] In modern manufacturing and multi-source data application scenarios, enterprises accumulate massive amounts of multimodal data, including customer demand data, product design drawings, specifications, production process parameters, equipment operation logs, maintenance records, etc. This data is scattered across different business systems, teams, and even individuals, creating a serious knowledge silo problem. Existing technologies generally face the following challenges when processing this data:

[0003] I. Severe knowledge silos hinder data collaboration

[0004] Data from various stages, including product development, R&D, production, and marketing, is scattered across different systems, resulting in extremely high communication costs and low efficiency for cross-departmental information retrieval, alignment, and collaboration. Existing solutions sometimes utilize large language models to construct multi-source knowledge graphs for cross-domain data queries; however, these solutions focus on data query methods and do not address the dynamic updating mechanism of the knowledge graph or the deployment of intelligent agents oriented towards business scenarios. Other solutions build knowledge graphs based on industrial big data but do not employ the RAG framework to process unstructured data; knowledge reasoning relies on a pre-defined rule base rather than the collaboration between the large model and the knowledge graph.

[0005] II. The model struggles to adapt to the temporal evolution of knowledge.

[0006] In the continuous evolution of knowledge, when the effective time of new knowledge exceeds the time coverage of existing knowledge, existing models struggle to adaptively expand their representational capabilities. Some existing incremental learning methods update model parameters by continuously fine-tuning the same model. However, in the process of fitting new data distributions, old knowledge is gradually overwritten and forgotten, leading to a catastrophic forgetting problem. For example, a related approach proposes a dynamic expert hybrid incremental graph learning method, adding a new expert network to each incoming data block in the GNN layer. However, its triggering condition is the data-driven event of "data block arrival," rather than a semantic state judgment based on knowledge, lacking the ability to perceive the effective time dimension of knowledge.

[0007] Third, the application of intelligent agents lacks a systematic coordination mechanism.

[0008] Existing knowledge graph systems mostly focus on intelligent applications in single business scenarios, lacking a multi-agent collaborative architecture that covers the entire business chain. Related solutions propose knowledge graph-guided LLM routers, which transform multi-agent question answering into knowledge graph-guided routing problems. However, their routing is based on semantic relevance matching, does not consider the temporal order constraints of business processes in scenarios such as manufacturing, and lacks a mechanism for chain-like transmission of intermediate inference results and context states between agents.

[0009] Furthermore, existing technologies have proposed a cognitive constraint of "read-search-verify-construct," but their search steps are based on semantic similarity retrieval and do not take the effective time attribute of knowledge as the core constraint condition for retrieval, thus failing to achieve time-sensitive and accurate knowledge retrieval.

[0010] In summary, none of the above solutions involve a dynamic update mechanism for knowledge graphs based on the effective time attribute. Although they involve incremental learning with dynamic expert hybridity, their triggering mechanism is "data block arrival" rather than semantic judgment based on the effective time coverage. Although they involve agent routing and reasoning, they do not take the effective time of knowledge as the core constraint for routing or retrieval. Existing technologies lack a unified platform that can integrate multi-source heterogeneous data, manage the effective time of knowledge, adapt to knowledge evolution, and support intelligent decision-making across the entire business chain. Summary of the Invention

[0011] This invention aims to address the aforementioned problems in the existing technology by providing a dynamic knowledge graph construction system and method based on multi-source data fusion. Specific objectives include:

[0012] First, it provides a knowledge graph construction system and method that can integrate multi-source heterogeneous data and manage the effective time of knowledge, so that each piece of knowledge carries the time interval information of its actual effectiveness in the real world, and supports accurate knowledge retrieval and historical tracing based on the time dimension.

[0013] Second, a dynamic update mechanism and method for knowledge graphs that can adapt to the evolution of knowledge over time is provided. When the effective time of newly added knowledge exceeds the coverage of existing knowledge, the model capacity is automatically expanded to fit the new data distribution, while avoiding catastrophic forgetting.

[0014] Third, it provides a multi-agent collaborative architecture and method covering multiple business scenarios, enabling intelligent routing and chain collaboration among agents, and supporting complex query processing across business processes.

[0015] Fourth, a continuous optimization mechanism and method based on implicit user feedback is provided. By collecting users' correction operations, the retrieval strategy and knowledge fusion quality are automatically optimized, forming a self-learning and self-optimizing closed-loop system.

[0016] To achieve the above objectives, the present invention adopts the following technical solution:

[0017] The first aspect of this invention provides a dynamic knowledge graph construction system based on multi-source data fusion, comprising:

[0018] The data layer integrates heterogeneous data from multiple sources. The knowledge layer receives data, performs entity recognition, relation extraction, and knowledge fusion to generate a knowledge graph, and labels each piece of knowledge with a valid time attribute, which represents the actual effective time interval of the knowledge. The technology layer includes a graph neural network, a retrieval enhancement generation framework, and an incremental learning engine. The incremental learning engine acquires the valid time attribute of each piece of knowledge. When the valid time endpoint of newly added knowledge is later than the latest valid endpoint of existing knowledge, it is determined that it exceeds the coverage of existing knowledge, and an expert network module is added to the graph neural network. The expert network module initializes new parameter nodes and updates them based on the gradient of the newly added knowledge to fit its data distribution. The application layer includes pre-sales intelligent agents, R&D intelligent agents, manufacturing intelligent agents, and service intelligent agents. The valid time attribute of the knowledge layer triggers the addition of the expert network module in the technology layer. The application layer collects user feedback and optimizes the fusion quality of the knowledge layer through the technology layer.

[0019] Furthermore, the knowledge layer breaks down multi-source heterogeneous data into atomic fact units, which are the smallest knowledge units containing a single fact statement. Knowledge is extracted independently from each atomic fact unit, and after sorting them according to the order of their effective time attributes, a binary merge algorithm based on the comparison of the endpoints of the effective time interval is used to merge and fuse the sorted results pairwise. The merged results are then written into the knowledge graph. During the merging process, for multiple relations of the same entity pair within the same time interval, the relation with the longest effective time interval is retained and the contained relation is discarded.

[0020] Furthermore, the graph neural network includes multiple expert network modules and a sparse gating network; during inference, the sparse gating network calculates the degree of overlap between the time coverage range modeled by each expert network module and the time range specified by the user query, and activates at least one expert network module with the highest degree of overlap to participate in the calculation.

[0021] Furthermore, the technical layer also includes a router. The router encodes the user query, the capability descriptions of each agent, and the entity types and relation types located in the knowledge graph into a heterogeneous graph structure. The nodes of the heterogeneous graph structure include query nodes, entity type nodes, relation type nodes, and agent nodes. The router performs reasoning on the heterogeneous graph structure through a graph neural network to predict the matching score between each agent and the user query, and determines the set of agents involved in the user query based on the matching score. When the set of agents contains multiple agents, the router triggers the multiple agents to process them sequentially according to the business process of pre-sales, R&D, manufacturing, and service. The previous agent transmits the output containing intermediate reasoning results and context state to the next agent through a standardized data interface.

[0022] Furthermore, each agent performs reasoning using a cognitive loop, which includes: a reading step, parsing the user's query intent and extracting the specified time range; a searching step, retrieving knowledge from the knowledge graph based on the parsing results, matching the effective time attributes of the knowledge with the time range, and retrieving only knowledge whose effective time intervals overlap with the time range; a verification step, performing consistency checks on the retrieved knowledge and excluding knowledge with contradictory effective time attributes; and a construction step, constructing a response result based on the verified knowledge and associating it with the original data source.

[0023] Furthermore, the knowledge layer stores the knowledge graph in a two-layer structure: the first layer is a temporal knowledge graph layer, which stores the relationships between entities using valid time intervals as edge labels; the second layer is a time index layer, which constructs an inverted index using valid time intervals as keys, with each valid time interval associated with a corresponding edge in the temporal knowledge graph layer; when the technology layer receives a user query containing a specified time range, it searches for all valid time intervals that overlap with the specified time range through the time index layer, and retrieves the knowledge represented by the corresponding edge from the temporal knowledge graph layer.

[0024] Furthermore, the application layer collects the user's correction operations on the agent's response results, identifies the text fragments that have been deleted or replaced by the user, performs semantic similarity matching between the text fragments and the knowledge stored in the knowledge graph, and marks the stored knowledge with semantic similarity exceeding a preset threshold as negative example signals; the technology layer adjusts the recall threshold of the retrieval enhancement generation framework according to the negative example signals, and the adjusted recall threshold is used to filter the candidate set of entity links in the subsequent knowledge fusion process, reducing the probability that the entity involved in the knowledge corresponding to the negative example signal will be linked again.

[0025] Furthermore, after the knowledge graph is updated, the knowledge layer searches for the valid time intervals involved in this update based on the time index layer, and only regenerates the knowledge summaries associated with the valid time intervals. The knowledge summaries are general descriptions of the knowledge within each valid time interval, while the knowledge summaries of the valid time intervals not involved remain unchanged.

[0026] The second aspect of this invention provides a method for constructing a dynamic knowledge graph based on multi-source data fusion, comprising the following steps: Step S1, integrating multi-source heterogeneous data through a data layer; Step S2, receiving data through a knowledge layer, performing entity recognition, relation extraction, and knowledge fusion to generate a knowledge graph, and labeling each piece of knowledge with an effective time attribute, wherein the effective time attribute represents the actual effective time interval of the knowledge; Step S3, providing technical support for the construction and updating of the knowledge graph through a graph neural network, a retrieval enhancement generation framework, and an incremental learning engine at the technical layer; obtaining the effective time attribute of each piece of knowledge, and determining that when the effective time endpoint of newly added knowledge is later than the latest effective endpoint of existing knowledge, it exceeds the coverage of existing knowledge, and adding an expert network module to the graph neural network, wherein the expert network module initializes new parameter nodes and updates based on the gradient of the newly added knowledge to fit its data distribution; Step S4, providing business decision-making services to users based on the knowledge graph through pre-sales intelligent agents, R&D intelligent agents, manufacturing intelligent agents, and service intelligent agents at the application layer.

[0027] Furthermore, the graph neural network in step S3 includes multiple expert network modules and a sparse gating network; during inference, the sparse gating network calculates the degree of overlap between the time coverage range modeled by each expert network module and the time range specified by the user query, and activates at least one expert network module with the highest degree of overlap to participate in the calculation; step S4 also includes: collecting the user's correction operations on the agent's response results, identifying the text fragments deleted or replaced by the user, performing semantic similarity matching between the text fragments and the knowledge stored in the knowledge graph, marking the stored knowledge with semantic similarity exceeding a preset threshold as negative example signals; adjusting the recall threshold of the retrieval enhancement generation framework according to the negative example signals, and using the adjusted recall threshold to filter the candidate set of entity links in the subsequent knowledge fusion process.

[0028] Compared with the prior art, the present invention has the following beneficial effects:

[0029] I. Achieving Precise Time Management of Knowledge. This invention labels each piece of knowledge with an effective time attribute, i.e., the time interval during which the knowledge is actually effective in the real world. It employs a two-layer time-series graph storage structure with the effective time interval as the edge label and the inverted index key. This enables the system to accurately answer queries containing time constraints, such as "What are the effective process standards in 2023?" This fundamentally solves the problem of traditional knowledge graph systems being unable to distinguish between the effective time and the entry time of knowledge, achieving precise time traceability of knowledge.

[0030] II. Achieving Adaptive Expansion of the Model to Knowledge Temporal Evolution. This invention uses "whether the effective time endpoint of newly added knowledge is later than the latest effective endpoint of existing knowledge" as a trigger condition, automatically adding expert network modules when the data distribution drifts across time domains. Compared to DyMoE's "adding an expert upon arrival of each data block," this invention's trigger mechanism is based on the semantic state judgment of knowledge rather than data-driven events, avoiding the waste of computational resources and overfitting risks caused by frequently adding expert networks. Simultaneously, by retaining existing expert networks and only adding experts to fit the new distribution, the catastrophic forgetting problem in continuous learning is fundamentally avoided.

[0031] Third, achieving accurate time-aware inference routing. In this invention, the sparse gating network selects the activated expert network based on the degree of overlap between the query time range and the expert network's time coverage range. Compared with DyMoE's routing mechanism based on input data features, this invention explicitly introduces time semantics into the routing decision, enabling the inference stage to accurately select the expert network most relevant to the query time, thereby improving the accuracy and efficiency of inference.

[0032] IV. Achieving Semantic-Temporal Cooperative Routing for Multi-Agents. In this invention, the router uses a heterogeneous graph neural network (GNN) to uniformly process queries, entity types, relationship types, and semantic relationships between agent nodes. After determining the set of agents, it triggers chained collaboration according to the business process sequence of pre-sales, R&D, manufacturing, and service. Compared with AgentRouter's routing based solely on semantic matching, this invention adds business temporal constraints, ensuring that the multi-agent collaboration results conform to the logical order of the actual business process.

[0033] V. Achieving Auditable and Time-Sensitive Agent Reasoning. In this invention, the agent performs reasoning using a "read-search-verify-construct" cognitive loop. The search step uses matching of valid time attributes as the core constraint for retrieval, while the verification step excludes knowledge with contradictory valid time attributes. Compared to the RAGA framework, this invention uses the time dimension as the core constraint of the cognitive loop rather than an optional filtering condition, ensuring the consistency and auditability of the agent's reasoning results in the time dimension.

[0034] VI. Achieving Continuous Closed-Loop Optimization Based on Implicit Feedback. This invention collects user correction actions on responses, i.e., deleting or replacing text fragments, and uses semantic similarity matching to locate corresponding knowledge nodes in the knowledge graph and mark them as negative example signals. These signals are then used to adjust the recall threshold and candidate set filtering of entity links within the RAG framework. Compared to traditional feedback optimization schemes that rely on explicit scoring, this invention collects feedback signals without requiring additional user actions, and the feedback directly impacts the optimization of entity link accuracy, forming a complete closed loop of "user correction → negative example marking → threshold adjustment → candidate set filtering → link quality improvement". Attached Figure Description

[0035] Figure 1 This is a schematic diagram of the overall four-layer hierarchical architecture of the dynamic knowledge graph construction system based on multi-source data fusion of the present invention.

[0036] Figure 2 This is a schematic diagram illustrating the knowledge graph construction and dynamic update process of multi-source data fusion in this invention. Detailed Implementation

[0037] Terminology Explanation

[0038] To facilitate understanding of this invention, the following descriptions of specific terms used in this application are provided:

[0039] Atomic Fact Unit: A unit of knowledge that is the smallest indivisible unit containing a single factual statement. An atomic fact unit expresses an independent and complete factual information. For example, "Product A malfunctioned in March 2024" is an atomic fact unit, while "Product A malfunctioned in March 2024, and the solution was to replace part C" contains two facts and should be split into two atomic fact units.

[0040] Effective Time Attribute: This refers to the time interval during which knowledge is actually effective in the real world, expressed in the form of "start time - end time", such as [2024-01-01, 2025-12-31]. The effective time attribute is different from the knowledge's creation time, entry time, or update time. The latter only indicate the time when the knowledge is recorded by the system, while the effective time attribute indicates the time period during which the knowledge is actually applicable in real-world business.

[0041] Cognitive loop: refers to the iterative reasoning process followed by an intelligent agent when processing user queries, including four steps: reading, searching, verifying, and constructing. Through this loop, the intelligent agent progressively transforms user queries into verifiable and traceable responses.

[0042] Binary merge algorithm: In this application, it specifically refers to a merge and fusion algorithm based on the comparison of endpoints of valid time intervals. This algorithm merges the knowledge extraction results that have been sorted by time pairwise, using the start and end points of the valid time interval as the comparison criteria, and resolves conflicts between multiple relations of the same entity within the same time interval.

[0043] Negative example signals: These are feedback signals identified through user correction actions, indicating that a piece of knowledge in the knowledge graph does not match the user's expectations. Negative example signals are used to adjust retrieval strategies and reduce the probability of the corresponding knowledge being incorrectly recalled or linked again.

[0044] The specific embodiments of the present invention will now be described in detail with reference to the accompanying drawings and examples. It should be understood that the specific embodiments described herein are for illustrative purposes only and are not intended to limit the scope of the invention.

[0045] Example 1: System Overall Architecture

[0046] like Figure 1 As shown, the dynamic knowledge graph construction system based on multi-source data fusion of the present invention includes a data layer, a knowledge layer, a technology layer, and an application layer.

[0047] The data layer integrates multi-source heterogeneous data from enterprises. Taking an imaging module manufacturing company as an example, the data integrated in the data layer includes a customer demand database, containing customer classification information, demand communication records, etc.; a product database, containing design drawings (PDF files), product specification documents (Word documents), patent documents, etc.; and a production database, containing BOM tables, equipment operation logs, work order records, etc.

[0048] The knowledge layer receives data from the data layer and performs entity recognition, relation extraction, and knowledge fusion through the dynamic knowledge graph construction module to generate a structured knowledge graph. The knowledge layer labels each piece of knowledge in the knowledge graph with a valid time attribute. For example, for the knowledge "Product model A has a resolution of 1080P" extracted from a product specification, if the effective date of the specification is January 1, 2024, the knowledge layer labels the valid time attribute of this knowledge as [2024-01-01, present]. Similarly, for the knowledge "Product A experienced a failure B in March 2024" extracted from maintenance records, the knowledge layer labels the valid time attribute of this knowledge as [2024-03-01, 2024-03-31], based on the failure occurrence time recorded in the maintenance records.

[0049] The technical layer includes a graph neural network, a retrieval enhancement and generation framework, and an incremental learning engine. In this embodiment, the graph neural network adopts a graph attention network architecture, with an initial number of four expert network modules, each corresponding to a different time coverage range. The RAG framework employs a vector database-based retrieval architecture, using the Sentence-BERT model to encode queries and knowledge fragments into vectors. The incremental learning engine uses a gradient update module implemented based on the PyTorch framework.

[0050] The application layer includes four intelligent agents: pre-sales intelligent agent, R&D intelligent agent, manufacturing intelligent agent, and service intelligent agent, which provide services to customers, R&D engineers, production managers, and after-sales service personnel, respectively.

[0051] Example 2: Atomic Fact Decomposition and Binary Merging

[0052] like Figure 2 As shown, the process of atomic fact decomposition and binary merging fusion in the knowledge layer is as follows:

[0053] First, the knowledge layer breaks down multi-source heterogeneous data into atomic fact units. For example, a product repair manual document might contain the following: "Product A frequently experienced fault B between January and June 2023, with the solution being to replace part C. Since July 2023, this fault has been completely resolved through process optimization."

[0054] The knowledge layer splits the document into two atomic units of fact:

[0055] Atomic Fact 1: "Product A experienced a malfunction B during the period [2023-01-01, 2023-06-30]".

[0056] Atomic Fact 2: "Fault B in Product A has been resolved since [July 1, 2023]".

[0057] After independently extracting knowledge from each atomic fact unit, the extraction results are sorted according to the order of the effective time attributes of each atomic fact unit, namely 2023-01-01 and 2023-07-01. Then, a binary merge algorithm based on the comparison of the endpoints of the effective time interval is used to merge the sorted results pairwise.

[0058] During the merging process, if multiple relations exist for the same entity pair, i.e., product A and fault B, within the same time interval, such as "fault B occurred" in atomic fact 1 and "fault B happened" extracted from another document, although the expressions are different, they have the same semantics. The binary merge algorithm retains the relation with the longest valid time interval, i.e., [2023-01-01, 2023-06-30], and discards the contained relations, such as [2023-03-01, 2023-04-30]. The merge results are sequentially written into the knowledge graph.

[0059] Example 3: Effective Time Attribute Triggers GNN Structure Expansion

[0060] In this embodiment, the process by which the effective temporal attribute of the technology layer triggers the GNN structure expansion is as follows:

[0061] The incremental learning engine continuously monitors the validity period of newly added knowledge in the knowledge layer. Assume the latest validity period for existing knowledge ends on December 31, 2025. When a company expands from the mobile phone module field to the automotive module field, a new knowledge entry is added to the knowledge layer: "The field of view of automotive module X is 120 degrees, valid from January 1, 2026."

[0062] The incremental learning engine obtains the effective time attribute of the newly added knowledge and finds that its effective time end point of December 31, 2026, is later than the latest effective end point of existing knowledge of December 31, 2025. Therefore, it determines that the newly added knowledge exceeds the coverage of existing knowledge. Consequently, the incremental learning engine adds an expert network module to the graph neural network, namely the fifth expert network module.

[0063] This expert network module initializes new parameter nodes and updates gradients based on newly added knowledge to fit the data distribution pattern of automotive module domain knowledge. Preferably, the initialization uses the Xavier uniform initialization method, with a learning rate set to 0.001 and 100 iterations. The parameters of the four existing expert network modules remain unchanged, thus preserving the modeling ability of mobile phone module domain knowledge and avoiding catastrophic forgetting.

[0064] Example 4: Time-Aware Sparse Gated Routing

[0065] In this embodiment, the time-aware sparse gating routing mechanism at the technical layer is as follows:

[0066] The graph neural network comprises multiple expert network modules and a sparse gating network. In this embodiment, there are 5 expert network modules, each modeling a different time range: expert 1 corresponds to 2020-2021, expert 2 corresponds to 2021-2022, expert 3 corresponds to 2022-2023, expert 4 corresponds to 2023-2024, and expert 5 corresponds to 2024 to the present.

[0067] When a user queries a specified time range of "the whole year of 2022", the sparse gating network calculates the degree of overlap between the time coverage range modeled by each expert network module and the query time range: the overlap between expert 1 and [2022-01-01, 2022-12-31] is 0%, the overlap between expert 2 and the query time range is 50%, the overlap between expert 3 and the query time range is 100%, the overlap between expert 4 and the query time range is 50%, and the overlap between expert 5 and the query time range is 0%.

[0068] The sparse gating network activates expert 3, which has the highest degree of overlap, to participate in the calculation. When the degree of overlap is the same, at least one expert network module with the highest degree of overlap is activated to participate in the calculation; in this embodiment, expert 3 is activated. If the query time range is "2021 to 2023", then the experts with the highest degree of overlap are expert 2, expert 3, and expert 4, and the sparse gating network activates these three expert modules to participate in the calculation together.

[0069] Example 5: Knowledge Graph-Guided Agent Routing and Chain Collaboration

[0070] In this embodiment, the workflow of the knowledge graph-guided intelligent agent router at the technical layer is as follows:

[0071] Suppose a user, i.e. a customer, enters the query: "We need a camera module for nighttime surveillance, which requires clear imaging under 0.1 lux illumination. Please provide a recommended solution and confirm its production feasibility."

[0072] After receiving the query, the router first locates the relevant entity types in the knowledge graph, including "camera module," "illuminance," and "imaging quality," as well as the relationship types, including "applicable," "meets requirements," and "manufacturable." Simultaneously, it obtains the capability descriptions of each agent: the pre-sales agent excels at product recommendation and solution design; the R&D agent excels at technical feasibility analysis; the manufacturing agent excels at production process evaluation; and the service agent excels at after-sales support.

[0073] The router encodes user queries, capability descriptions of each agent, and the entity and relation types located into a heterogeneous graph structure. The nodes of the heterogeneous graph structure include query node Q, entity type nodes "camera module", "illuminance", and "imaging quality", relation type nodes "suitable", "meets requirements", and "manufacturable", and agent nodes "pre-sales agent", "R&D agent", "manufacturing agent", and "service agent".

[0074] The router uses a graph neural network to reason about the heterogeneous graph structure and predict the matching score between each agent and the user query: 0.92 for pre-sales agent, 0.78 for R&D agent, 0.65 for manufacturing agent, and 0.12 for service agent.

[0075] Based on the matching scores, the router determines that the set of agents involved in the user query is {Pre-sales Agent, R&D Agent, Manufacturing Agent}, and these agents' matching scores exceed the threshold of 0.5. Since the set of agents contains multiple agents, the router triggers these three agents to process sequentially according to the business process order of pre-sales, R&D, and manufacturing.

[0076] The pre-sales agent processes the data first, retrieving matching product models based on a knowledge graph and generating an initial recommendation. It then transmits the intermediate inference results and context state to the R&D agent via a standardized data interface in JSON format, containing a list of recommended products, technical parameters, and the inference process. Upon receiving this information, the R&D agent performs a technical feasibility analysis based on the pre-sales agent's recommendations, confirming that the product parameters meet the 0.1 lux illuminance requirement, and transmits the analysis results to the manufacturing agent. The manufacturing agent evaluates the feasibility of the production process based on the outputs of the first two agents, ultimately generating a complete response that includes product recommendations, technical confirmation, and production assessment.

[0077] Example 6: Cognitive Loop Reasoning of an Intelligent Agent

[0078] In this embodiment, each agent performs reasoning using a cognitive loop. Taking the pre-sales agent processing the aforementioned customer query as an example:

[0079] Reading steps: The pre-sales agent parses the user's query intent, identifies the core requirement as "camera module for nighttime surveillance," and extracts the specified time range. If the user does not explicitly specify a time range, the current time is used as the time constraint for the search by default.

[0080] Search Steps: The pre-sales agent retrieves knowledge from the knowledge graph based on the parsing results, matching the valid time attribute of the knowledge with the current time range, and only retrieving knowledge whose valid time interval overlaps with the current time. Assuming the current year is 2026, the knowledge "Product Model X is suitable for nighttime monitoring, valid from [2023-01-01, 2025-12-31]" is excluded from the search results because its valid end date is earlier than 2026; the knowledge "Product Model Y is suitable for nighttime monitoring, valid from [2024-01-01, present]" is included in the search results because its valid time interval overlaps with 2026.

[0081] Verification Steps: The pre-sales agent performs a consistency check on the retrieved knowledge. For example, if two contradictory pieces of knowledge are retrieved—"Product Y has a resolution of 1080P" with a valid time [2024-01-01, 2025-06-30], and "Product Y has a resolution of 4K" with a valid time [2025-07-01, present], and the current query time is 2026—the pre-sales agent will exclude the older knowledge "1080P" which contradicts the current time and retain the valid new knowledge "4K".

[0082] Construction steps: The pre-sales agent constructs the response result based on the verified knowledge, recommends product Y and associates it with the original data source, namely a specific page number of the product specification PDF.

[0083] Example 7: Two-layer time sequence graph storage and time index retrieval

[0084] In this embodiment, the knowledge layer stores the knowledge graph in a two-layer structure:

[0085] The first layer is a temporal knowledge graph layer, storing the relationships between entities using valid time intervals as edge labels. For example, edge 1 represents the relationship "resolution" between entity "product Y" and entity "resolution value," with a value of 1080P and a valid time interval of [2024-01-01, 2025-06-30]; edge 2 represents the relationship "resolution" between entity "product Y" and entity "resolution value," with a value of 4K and a valid time interval of [2025-07-01, present]. Multiple relationships between the same entity pair, i.e., product Y and resolution value, at different times are stored separately using different edges, without overlapping.

[0086] The second layer is the time index layer, which constructs an inverted index using the valid time interval as the key. For example, [2024-01-01, 2025-06-30] corresponds to edge 1, edge 5, edge 12, etc., and [2025-07-01, present] corresponds to edge 2, edge 8, edge 15, etc.

[0087] When the technical layer receives a user query containing a specified time range, such as "query the resolution of product Y for the entire year of 2025", it searches through the time index layer for all valid time intervals that overlap with [2025-01-01, 2025-12-31], including [2024-01-01, 2025-06-30] and [2025-07-01, present], both of which partially overlap. Then, it retrieves the knowledge represented by the corresponding edges from the time-series knowledge graph layer, i.e., edge 1 corresponds to 1080P, covering January to June 2025, and edge 2 corresponds to 4K, covering July to December 2025, generating a complete response containing time segmentation information.

[0088] Example 8: Closed-loop optimization based on implicit user feedback

[0089] In this embodiment, the closed-loop optimization process based on implicit user feedback is as follows:

[0090] Taking R&D Agent as an example, suppose an R&D engineer enters the query: "What are the main failure modes of product A in 2023?" The R&D Agent, based on the knowledge graph, replies: "The main failure modes of product A in 2023 are failure B and failure C."

[0091] The engineer discovered that the reply contained "fault C," but based on his experience, fault C had not yet appeared in 2023 and would actually appear in 2024, so he removed "fault C" from the reply.

[0092] The application layer collects the correction operation data and identifies the deleted text fragment "Fault C". It performs semantic similarity matching between "Fault C" and the knowledge stored in the knowledge graph, calculating cosine similarity using the Sentence-BERT model with a threshold of 0.85. It finds that the semantic similarity between "Fault C" and the triple "Product A, Fault Mode, Fault C, Effective Time [2024-01-01, Present]" in the knowledge graph is 0.92, exceeding the preset threshold of 0.85 (the preset threshold is determined through F1-score tuning on a validation set, typically between 0.8 and 0.9). Therefore, this triple is marked as a negative example signal.

[0093] Based on this negative example signal, the technical layer adjusted the recall threshold of the RAG framework from 0.75 to 0.80, increasing the difficulty for subsequent searches to recall knowledge semantically similar to this negative example signal. The adjusted recall threshold is used to filter the candidate set of entity links during subsequent knowledge fusion, reducing the probability that the entity corresponding to "fault C" will be incorrectly linked to related queries in 2023 again.

[0094] Example 9: Incremental Summary Update Based on Time Interval Index

[0095] In this embodiment, the incremental summary update process of the knowledge layer after the knowledge graph is updated is as follows:

[0096] Suppose that the knowledge graph adds a new knowledge entry on March 15, 2026: "Product Z was released on March 10, 2026, with a resolution of 8K and an effective period of [2026-03-10, to date]".

[0097] The knowledge layer uses the time index layer to find the valid time intervals involved in this update. Since the valid time interval for newly added knowledge is [2026-03-10, present], the time index layer locates the index entry corresponding to this interval. The knowledge layer only regenerates the knowledge summary associated with this valid time interval. The knowledge summary is a general description of the knowledge within each valid time interval. For example, the knowledge summary for the interval [2026-03-10, present] is "Related information for product Z added on or after March 10, 2026". Valid time intervals not covered, such as [2025-01-01, 2025-12-31] and [2024-01-01, 2024-12-31], retain their knowledge summaries.

[0098] This incremental update method avoids the high computational overhead of full summary reconstruction, enabling the system to support high-frequency knowledge updates.

[0099] Example 10: Complete Business Process Example

[0100] The following example, using the complete business process of an image module manufacturing company, illustrates the overall working process of the system of this invention:

[0101] In the pre-sales stage: The customer submits a requirement through a pre-sales agent: "We need a 12-megapixel camera module for industrial inspection, with a size smaller than 10mm×10mm×10mm." The pre-sales agent processes this query using a cognitive loop: The reading step extracts key requirements such as "12-megapixel," "smaller than 10mm×10mm×10mm," and "industrial inspection," along with the current time range; the search step matches the current time with the valid time attributes of knowledge, retrieving product models that meet the conditions from the knowledge graph; the verification step excludes products that are discontinued or whose specifications have changed; and the construction step generates a recommended solution and associates it with the original data source, i.e., the product specification sheet. The pre-sales agent then transmits the recommended results to the R&D agent through a standardized interface.

[0102] R&D phase: The R&D agent receives the recommendation results from the pre-sales agent, combines patent documents, technical documents and design cases in the knowledge graph to design technical solutions, confirms that the technical parameters of the selected product model meet customer needs, and evaluates the feasibility of customized development.

[0103] Manufacturing stage: The manufacturing agent receives the analysis results from the R&D agent, assesses the production feasibility of the product based on the production process parameters and equipment operation logs in the knowledge graph, predicts possible yield problems, and provides process optimization suggestions.

[0104] Service Phase: After product delivery, the service agent provides intelligent operation and maintenance and after-sales service support to customers based on maintenance records and fault solutions in the knowledge graph. Customer usage feedback and maintenance records serve as new data sources flowing back to the data layer, triggering incremental updates in the knowledge layer and closed-loop optimization in the technology layer.

[0105] Data Compliance Statement

[0106] The collection of user correction data for agent responses in this application is limited to data on the operational behavior of internal employees during system use, including text fragments deleted or replaced by users. All collected data has been anonymized and is used solely for system optimization purposes. It does not involve personal privacy information and complies with the relevant provisions of the Personal Information Protection Law of the People's Republic of China.

Claims

1. A dynamic knowledge graph construction system based on multi-source data fusion, characterized in that, include: The data layer is used to integrate heterogeneous data from multiple sources. The knowledge layer is used to receive data, perform entity recognition, relation extraction, and knowledge fusion to generate a knowledge graph, and to label each piece of knowledge with a valid time attribute, which represents the time interval during which the knowledge is actually effective. The technical layer includes a graph neural network, a retrieval enhancement generation framework, and an incremental learning engine. The incremental learning engine acquires the effective time attribute of each knowledge. When the effective time endpoint of the newly added knowledge is later than the latest effective endpoint of the existing knowledge, it is determined that it exceeds the coverage of the existing knowledge. An expert network module is added to the graph neural network. The expert network module initializes new parameter nodes and updates them based on the gradient of the newly added knowledge to fit its data distribution. The application layer includes pre-sales intelligent agents, R&D intelligent agents, manufacturing intelligent agents, and service intelligent agents; The effective time attribute of the knowledge layer triggers the addition of an expert network module in the technology layer, and the application layer collects user feedback and optimizes the fusion quality of the knowledge layer through the technology layer.

2. The dynamic knowledge graph construction system based on multi-source data fusion according to claim 1, characterized in that, The knowledge layer breaks down the multi-source heterogeneous data into atomic fact units, each of which is the smallest knowledge unit containing a single fact statement. Knowledge is extracted independently from each atomic fact unit, and the atomic fact units are sorted according to their effective time attributes. The sorting results are then merged pairwise using a binary merge algorithm based on the comparison of endpoints of the effective time interval, and the merged results are written into the knowledge graph in sequence. During the merging process, for the same entity pair with multiple relations within the same time interval, the relation with the longest valid time interval is retained and the contained relation is discarded.

3. The dynamic knowledge graph construction system based on multi-source data fusion according to claim 1, characterized in that, The graph neural network comprises multiple expert network modules and a sparse gating network; During inference, the sparse gating network calculates the degree of overlap between the time coverage range modeled by each expert network module and the time range specified by the user query, and activates at least one expert network module with the highest degree of overlap to participate in the calculation.

4. The dynamic knowledge graph construction system based on multi-source data fusion according to claim 1, characterized in that, The technical layer also includes a router, which encodes the user query, the capability descriptions of each agent, and the entity types and relation types located in the knowledge graph into a heterogeneous graph structure. The nodes of the heterogeneous graph structure include query nodes, entity type nodes, relation type nodes, and agent nodes. The router performs reasoning on the heterogeneous graph structure through a graph neural network to predict the matching score between each agent and the user query, and determines the set of agents involved in the user query based on the matching score. When the set of agents contains multiple agents, the router triggers the multiple agents to process them sequentially according to the business process of pre-sales, R&D, manufacturing, and service. The previous agent transmits the output containing intermediate reasoning results and context state to the next agent through a standardized data interface.

5. The dynamic knowledge graph construction system based on multi-source data fusion according to claim 1, characterized in that, Each agent reasones using a cognitive loop, which includes: The reading process involves parsing the user's query intent and extracting the specified time range. The search step involves retrieving knowledge from the knowledge graph based on the parsing results, matching the effective time attributes of the knowledge with the time range, and retrieving only knowledge whose effective time interval overlaps with the time range. The verification step involves checking the consistency of the retrieved knowledge and excluding knowledge with contradictory valid time attributes. The construction steps involve constructing response results based on the verified knowledge and linking them to the original data source.

6. The dynamic knowledge graph construction system based on multi-source data fusion according to claim 1, characterized in that, The knowledge layer stores the knowledge graph in a two-layer structure: the first layer is a temporal knowledge graph layer, which stores the relationships between entities using effective time intervals as edge labels; The second layer is the time index layer, which constructs an inverted index using valid time intervals as keys. Each valid time interval is associated with a corresponding edge in the time-series knowledge graph layer. When the technology layer receives a user query containing a specified time range, it uses the time index layer to find all valid time intervals that overlap with the specified time range and retrieves the knowledge represented by the corresponding edge from the time-series knowledge graph layer.

7. The dynamic knowledge graph construction system based on multi-source data fusion according to claim 1, characterized in that, The application layer collects user correction operations on the agent's response results, identifies text fragments that have been deleted or replaced by the user, performs semantic similarity matching between the text fragments and the knowledge stored in the knowledge graph, and marks the stored knowledge with semantic similarity exceeding a preset threshold as negative example signals; the technology layer adjusts the recall threshold of the retrieval enhancement generation framework according to the negative example signals, and the adjusted recall threshold is used to filter the candidate set of entity links in the subsequent knowledge fusion process, reducing the probability that the entities involved in the knowledge corresponding to the negative example signals will be linked again.

8. The dynamic knowledge graph construction system based on multi-source data fusion according to claim 6, characterized in that, After the knowledge graph is updated, the knowledge layer searches for the valid time intervals involved in this update based on the time index layer, and only regenerates the knowledge summaries associated with the valid time intervals. The knowledge summaries are general descriptions of the knowledge within each valid time interval, while the knowledge summaries of valid time intervals not involved remain unchanged.

9. A method for constructing a dynamic knowledge graph through multi-source data fusion, characterized in that, Includes the following steps: Step S1: Integrate multi-source heterogeneous data through the data layer; Step S2: Receive data through the knowledge layer, perform entity recognition, relation extraction, and knowledge fusion to generate a knowledge graph, and label each piece of knowledge with an effective time attribute, wherein the effective time attribute represents the time interval during which the knowledge is actually effective; Step S3: Technical support is provided for the construction and updating of the knowledge graph through the graph neural network, retrieval enhancement generation framework and incremental learning engine of the technical layer; the effective time attribute of each knowledge is obtained; when the effective time end point of the new knowledge is later than the latest effective end point of the existing knowledge, it is determined that it exceeds the coverage of the existing knowledge, and an expert network module is added to the graph neural network. The expert network module initializes new parameter nodes and updates them based on the gradient of the new knowledge to fit its data distribution. Step S4: Based on the knowledge graph, the pre-sales intelligent agent, R&D intelligent agent, manufacturing intelligent agent and service intelligent agent of the application layer provide business decision-making services to users.

10. The method according to claim 9, characterized in that, The graph neural network in step S3 includes multiple expert network modules and a sparse gating network. During inference, the sparse gating network calculates the degree of overlap between the time coverage range modeled by each expert network module and the time range specified by the user query, and activates at least one expert network module with the highest degree of overlap to participate in the calculation. Step S4 further includes: collecting user correction operations on the agent's response results, identifying text fragments deleted or replaced by the user, performing semantic similarity matching between the text fragments and the knowledge stored in the knowledge graph, marking the stored knowledge with semantic similarity exceeding a preset threshold as negative example signals; adjusting the recall threshold of the retrieval enhancement generation framework according to the negative example signals, and using the adjusted recall threshold to filter the candidate set of entity links in the subsequent knowledge fusion process.

Images