A multi-level knowledge management method based on semantic request driving

Abstract

The application belongs to the technical field of distributed knowledge graph, and particularly relates to a multi-level knowledge management method based on semantic request driving, which comprises the following steps: for a knowledge graph multi-node system, an agent edge node maintains a change mark and a semantic vector index of a local knowledge base; for a change operation of knowledge resources, a semantic entropy of knowledge change is defined to quantify the semantic impact degree of a single change on a distributed system; the semantic entropy is used as a unified driving signal throughout the three processes of "reporting-requesting-issuing"; in the reporting stage, the semantic entropy determines whether to trigger reporting and selects a routing strategy; in the requesting stage, the semantic entropy influences the calculation of adaptive TTL and the cache strategy; in the issuing stage, the semantic entropy accumulation triggers active pushing; the method can use a unified semantic protocol throughout the whole life cycle of knowledge, and realize multi-level closed-loop scheduling knowledge management.

Description

Technical Field

[0001] This invention belongs to the field of distributed knowledge graph technology, specifically involving a multi-level knowledge management method based on semantic request-driven approach, which is applicable to the full lifecycle management of knowledge in large-scale distributed multi-node network scenarios such as the Internet of Things and the Industrial Internet. Background Technology

[0002] With the rapid development of large-scale distributed multi-node networks such as the Internet of Things (IoT) and the Industrial Internet, traditional centralized knowledge management methods are struggling to meet the demands of massive nodes for real-time knowledge sharing and efficient flow due to issues such as single-point-of-failure risks, severe data silos, and low collaboration efficiency. To address this, distributed knowledge graph technology, by distributing knowledge data across multiple nodes, effectively improves the system's scalability and fault tolerance, and has demonstrated significant application value in fields such as autonomous driving and smart cities.

[0003] Among existing distributed knowledge management methods, file synchronization-based methods suffer from low transmission efficiency and poor real-time performance; database replication-based methods only focus on data-level consistency and cannot support semantic-level differentiated processing; event-driven methods based on message queues offer good real-time performance, but lack a unified semantic annotation mechanism, resulting in each process running independently and failing to form a closed-loop management system. Methods that implement knowledge flow through unified message formats (such as RESTful API-based interface calls and message publish / subscribe based on message queue telemetry transmission) only transmit the data itself, without carrying semantic information about the resources. This prevents the central service and proxy layer from performing differentiated processing based on semantics, causing all requests to follow a uniform processing path, resulting in low processing efficiency and poor flexibility.

[0004] Furthermore, traditional distributed knowledge graph systems, such as Apache Jena and Neo4j, often develop knowledge storage, querying, and synchronization as independent modules, lacking coordinated scheduling based on a unified semantic protocol between these processes. In practical applications, facing complex scenarios such as high-concurrency updates of knowledge data, dynamic additions and removals of nodes, and changing network environments, existing systems still have significant technical shortcomings in areas such as semantic differentiation processing capabilities, efficiency of coordinated scheduling between multi-level nodes, and the balance between data consistency and system performance.

[0005] In knowledge graph-driven multi-node network systems, knowledge data continuously flows between network nodes, proxy services, central services, and autonomous services. Knowledge management encompasses three core processes: knowledge reporting, requesting, and distributing. Existing technologies have the following shortcomings: First, the knowledge reporting process lacks unified semantic annotation and differentiated processing capabilities. Currently, when edge nodes report knowledge change data to the central service, they typically only transmit the raw data content, without including necessary semantic description information (such as resource type, operational semantics, and scope of impact) in the protocol. This prevents the central service from implementing differentiated processing strategies based on the inherent semantics of the reported data, forcing all reported data to follow a homogeneous processing logic, severely restricting the system's processing efficiency and flexibility.

[0006] Second, the knowledge request process lacks an efficient multi-level scheduling mechanism. When edge nodes request knowledge data, existing methods typically use a fixed hierarchical query order (e.g., "local → proxy → autonomous → center"), without dynamically selecting the optimal query level based on real-time costs (latency, bandwidth, freshness, etc.). Even when the autonomous service holds valid but outdated data, it still prioritizes querying the autonomous service and returns expired data; when the autonomous service's data is missing, it still queries level by level in a fixed order, increasing latency. Furthermore, existing methods lack predictive acquisition based on semantic similarity, failing to reduce the number of subsequent requests.

[0007] Third, knowledge delivery lacks awareness of the cumulative impact of changes and guarantees of causal consistency. When the central service proactively delivers knowledge to edge nodes, existing methods typically employ full delivery or fixed-period delivery, failing to perceive the accumulation of semantic entropy from knowledge changes (i.e., the cumulative degree of the impact of changes) or the vector clock version of edge nodes (i.e., the degree of causal lag). Full delivery involves large amounts of data and is time-consuming; fixed-period delivery may miss critical changes or repeatedly deliver unchanged data, wasting bandwidth. Furthermore, because it is impossible to determine the current knowledge version held by a node, differentiated synchronization cannot be quickly performed on newly activated nodes, resulting in nodes remaining in a state of knowledge deficiency for extended periods, impacting business continuity. Summary of the Invention

[0008] In view of this, the present invention provides a multi-level knowledge management method based on semantic request-driven approach, which can use a unified semantic protocol to run through the entire knowledge lifecycle and realize multi-level closed-loop scheduling of knowledge management.

[0009] The technical solution for implementing the present invention is as follows: In a first aspect, the present invention provides a multi-level knowledge management method based on semantic request-driven approaches, comprising: Semantic entropy definition: For a multi-node knowledge graph system, the edge nodes of an agent maintain change markers and semantic vector indexes for the local knowledge base. For a single change operation of a knowledge resource, the semantic entropy of the knowledge change is defined to quantify the degree of semantic impact of a single change on the distributed system. Knowledge reporting: When an agent's edge node locally changes knowledge data, it constructs a semantic impact degree based on the semantic entropy of the knowledge change. It integrates three dimensions—the historical change frequency of knowledge resources, the semantic entropy of the current knowledge change, and the global dependency degree—into a unified triggering quantification indicator to replace the traditional reporting triggering logic that only relies on data volume or timestamps. When the semantic impact degree exceeds the adaptive reporting threshold, the agent's edge node is triggered to report the locally changed knowledge data to the central service. Knowledge Request: When an edge node of an agent requests knowledge data from the central service, the agent service layer implements an adaptive multi-level scheduling decision based on a cost function according to the knowledge semantic descriptor, and uses semantic entropy to calculate the adaptive TTL to perform cache verification. Knowledge distribution: The central service maintains a semantic entropy accumulation value for each knowledge resource and only triggers active distribution when the accumulation value crosses the distribution threshold; by calculating the overall deviation between the knowledge held by the computing nodes and the central full knowledge in the semantic vector space, it dynamically selects full synchronization, incremental synchronization or differential synchronization.

[0010] Semantic entropy serves as a unified driving signal throughout the three processes of knowledge reporting, knowledge request, and knowledge distribution.

[0011] Optionally, the semantic entropy of knowledge change described in this invention is calculated from the influence probability. During the construction of the influence probability, the dependency weights of nodes on resources in the knowledge graph topology and the real-time node activity status are integrated, so that the entropy value can dynamically reflect the semantic influence distribution of the change operation among active downstream nodes.

[0012] Optionally, this invention relates to knowledge resources. At any moment A change operation The semantic entropy of knowledge change is defined as:

[0013] in, For change operation For the first The probability of influence of each downstream consumption node This represents the total number of downstream consumption nodes.

[0014] Optionally, the influence probability described in this invention for:

[0015] in, It is a very small number. For nodes Resources Dependency weights For a moment The set of nodes that are active, 1[ ] is an indicator function.

[0016] Optionally, the semantic influence degree described in this invention for:

[0017] in, Representing resources Normalized value of historical change frequency Representing resources The degree of dependence, , which is a magnification factor.

[0018] Optionally, the proxy service layer of this invention implements an adaptive multi-level scheduling decision based on a cost function according to the knowledge semantic descriptor, specifically as follows: Calculate the scheduling cost at each level. for:

[0019] in, For the first The expected response delay cost of a service of the highest level. For from the first Level to acquire resources The estimated bandwidth consumption cost; For the first Level holding resources The cost of freshness; These are the weighting coefficients; The optimal query level is selected as follows:

[0020] Among them, level 1 is the proxy service layer cache, level 2 is the autonomous service layer, and level 3 is the central service layer.

[0021] Optionally, the process of using semantic entropy to calculate adaptive TTL and perform cache verification as described in this invention is as follows: The adaptive TTL is calculated as follows:

[0022] in, Based on the expiration time, This is the semantic entropy decay coefficient. This is the volatility attenuation coefficient. Representing knowledge resources The semantic entropy statistical feature value, For resources The total number of operations within a time window, when hour, , For resources The normalized value of the historical change frequency.

[0023] Optionally, when multiple resources need to be distributed simultaneously, the present invention arranges them in descending order of their accumulated semantic entropy values, prioritizing the distribution of resources with high semantic entropy, ensuring that high-impact changes reach each node first.

[0024] Optionally, the present invention computes the degree of overall deviation between the knowledge held by the nodes and the total knowledge of the center in the semantic vector space. (Semantic consistency deviation) is:

[0025] in, and They are nodes and resources in the center service semantic vectors, For semantic similarity threshold, Represents a node The knowledge set held This represents the complete set of knowledge held by the central service. Represents cosine semantic similarity. ; Dynamically select synchronization strategy based on semantic consistency deviation:

[0026] in, To achieve full synchronization, all knowledge related to the edge nodes of the intelligent agent is extracted and distributed. For incremental synchronization, causal increments are extracted based on vector clocks and then distributed. For differential synchronization, only resources whose semantic entropy accumulation exceeds a threshold are synchronized, minimizing the amount of data transmitted. This represents the full synchronization threshold. When the semantic deviation reaches this value or above, the incremental / differential synchronization is insufficient to repair it, and a full reconstruction is required. This represents the incremental synchronization threshold. When the semantic deviation is below this value, the node is basically consistent with the center and only differential repair is needed.

[0027] Optionally, the present invention also includes task-driven targeted knowledge delivery, wherein the knowledge semantic descriptor KSD also carries a task identifier, priority, and deadline.

[0028] Beneficial effects First, semantic-driven unified management enables differentiated closed-loop scheduling across processes. This invention proposes the concept of knowledge semantic entropy, elevating the impact of knowledge changes from a qualitative description to a calculable quantitative indicator. Existing systems such as Apache Jena and Neo4j can only perform qualitative impact analysis and cannot provide calculable entropy measurements. This invention uses semantic entropy as a unified driving signal throughout the three processes of "reporting-requesting-deploying": in the reporting stage, semantic entropy determines whether to trigger reporting and which routing strategy to select; in the request stage, semantic entropy affects the calculation of adaptive TTL and caching strategies; in the deployment stage, the accumulation of semantic entropy triggers proactive push. This enables the proxy layer, which was originally stateless and only executed fixed forwarding logic, to perform stateful differentiated closed-loop scheduling based on semantic quantification, a capability that existing general-purpose transport protocols and knowledge graph systems do not possess.

[0029] Second, semantic entropy accumulation triggers delivery and three-level synchronization, enabling precise on-demand push. This invention's proactive knowledge delivery abandons the traditional extensive mode of full delivery or fixed-period delivery. It maintains the accumulated semantic entropy value of each knowledge resource through a central service, triggering proactive delivery only when the accumulation exceeds a threshold. This fundamentally avoids the duplicate push of unchanged resources and the omission of critical changes during fixed-period delivery. For node activation scenarios, this invention constructs a synchronization strategy decision mechanism based on semantic consistency deviation—dynamically selecting full synchronization, incremental synchronization, or differential synchronization based on the overall deviation between the node and the center in the semantic vector space. Differential synchronization further filters high-impact resources for precise delivery through a semantic entropy accumulation threshold, minimizing transmission volume and enabling newly activated nodes to quickly restore business continuity.

[0030] Third, adaptive multi-level scheduling based on a cost function significantly reduces read latency. This invention constructs an adaptive scheduling mechanism based on a multi-objective cost function, replacing the fixed sequential query pattern of "local → proxy → autonomous → center" in existing methods. This cost function comprehensively considers the real-time status of three dimensions: response latency, bandwidth consumption, and semantic freshness, dynamically selecting the optimal query level. Simultaneously, the proxy layer performs predictive prefetching based on semantic similarity, enabling subsequent related queries to directly retrieve data from the local machine or cache. Coupled with semantic entropy-driven adaptive TTL and an active cache invalidation broadcast mechanism after changes in central service data, this systematically solves the latency jitter, cache penetration, and data inconsistency problems of hot data queries in distributed knowledge graphs, significantly reducing the frequency and bandwidth costs of cross-network access. Attached Figure Description

[0031] To more clearly illustrate the technical solutions of the embodiments of the present invention, 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.

[0032] Figure 1 This is the overall system architecture of the present invention; Figure 2 This is a flowchart illustrating the knowledge reporting process of the method of this invention. Figure 3 This is a flowchart illustrating the knowledge request process of the method of this invention. Figure 4 is a flowchart of the proactive knowledge dissemination process of the method of the present invention. Detailed Implementation

[0033] The embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0034] It should be noted that, in the absence of conflict, the following embodiments and features can be combined with each other; and, based on the embodiments of this disclosure, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this disclosure.

[0035] It should be noted that various aspects of embodiments within the scope of the appended claims are described below. It will be apparent that the aspects described herein can be embodied in a wide variety of forms, and any particular structure and / or function described herein is merely illustrative. Based on this disclosure, those skilled in the art will understand that one aspect described herein can be implemented independently of any other aspect, and two or more of these aspects can be combined in various ways. For example, any number of aspects set forth herein can be used to implement the device and / or practice the method. Additionally, this device and / or method can be implemented using structures and / or functionalities other than one or more of the aspects set forth herein.

[0036] like Figure 1 As shown, the overall system architecture and semantic request protocol, and the functions of each layer are as follows: ①Agent Layer: Running on each physical node, responsible for collecting data, maintaining local knowledge subsets, initiating reports and requests, receiving and merging data, and each agent maintains a local knowledge base, semantic vector index, and vector clock.

[0037] ② Proxy Service Layer: Deployed between the agent layer and the central service layer, it provides functions such as semantic entropy calculation, adaptive caching, cost function scheduling, flight deduplication, semantic prefetching and asynchronous backfilling, and is the core forwarding layer of closed-loop control.

[0038] ③ Autonomous Service Layer: A lightweight knowledge service located close to the proxy service layer, caching a subset of commonly used knowledge from the proxy service layer and providing low-latency local reading.

[0039] ④ Central Service Layer: The full knowledge storage center performs global operations such as knowledge persistence, anomaly detection, semantic entropy accumulation monitoring, incremental extraction, and distribution.

[0040] The knowledge semantic descriptor protocol designed in this invention is shown in the table below: Table 1 Knowledge Semantic Descriptor Protocol Table

[0041] The semantic vector, semantic_vector, is generated by a lightweight graph encoder (such as GraphSAGE) using the type, attributes, and description of knowledge items, and is used to measure the semantic similarity between knowledge items.

[0042] Knowledge reporting process like Figure 2 As shown, the edge nodes of the intelligent agent report locally changed knowledge data to the central service. The specific steps are as follows: Step 101, semantically aware change detection The agent's edge nodes maintain change tags and semantic vector indexes for their local knowledge base. For knowledge resources... At any moment A change operation Its knowledge change semantic entropy is defined as:

[0043] in, For change operation For the first The probability of the impact of each downstream consumption node ( To ensure the fraction is a very small number (to prevent the denominator from being 0), For nodes Resources The dependency weights are calculated using the weights of the relation edges in the knowledge graph. For a moment The set of nodes that are active, 1[ ] is an indicator function. This indicates that the change only affects a very small number of nodes; This means the change affects all active nodes. This represents the total number of active nodes.

[0044] When knowledge data is updated (attribute values ​​change, relation edges are added or deleted), the semantic entropy of the knowledge change is calculated. Calculate the semantic impact of the change:

[0045] in, Representing resources Normalized value of historical change frequency This represents the semantic entropy of the current change operation. Representing resources The degree of dependence, , which is the dependency amplification factor. When semantic influence... Exceeding the adaptive reporting threshold Knowledge reporting is triggered in a timely manner. In this embodiment of the invention, the adaptive reporting threshold... The threshold is dynamic and its value is determined based on the exponentially weighted moving average of the historical reporting frequency of the edge nodes of the agent. The initial default value is set to 0.5, and the value range is limited to [0.2, 0.9]. When the node is in a state of limited network bandwidth, the threshold is dynamically adjusted to ensure that the reporting frequency is lower than the network carrying capacity limit.

[0046] Step 102, Construction of Knowledge Semantic Descriptor Construct a reporting request, carrying a complete Knowledge Semantic Descriptor (KSD), including resource_uri (resource identifier), semantic_vector (semantic vector), operation_semantics (operation type), semantic_entropy (semantic entropy value), consistency_stamp (vector clock), source_topology (topology location), and correction_policy (correction flag).

[0047] Step 103, Differentiated routing based on semantic entropy After receiving the reporting request, the proxy service selects a differentiated routing strategy based on the semantic_entropy value in KSD.

[0048]

[0049] in, For low-impact changes, write to the target shard node via a consistent hash ring; To mitigate the impact of changes, write the changes to the primary node and synchronize them to the backup nodes; For high-impact changes, the broadcast is written to all relevant segments. Among them, and In this embodiment, the preset semantic entropy classification threshold is used. Values , Values .

[0050] Step 104: Central Service Processing and Anomaly Detection After receiving data, the central service executes the corresponding operation based on the operation type (operation_semantics). If the error correction flag (correction_policy=true) and the operation type is create / update, an anomaly detection algorithm is invoked to detect attribute values ​​in the reported data. Outliers are then filled in with implied values, which are then overwritten and stored. For example, an anomaly detection model can be trained on historical data using the Isolation Forest algorithm. This model assigns anomaly scores to reported numerical attribute values ​​(the closer the score is to 1, the higher the degree of anomaly). When the anomaly score exceeds a threshold, the attribute value is marked as an anomaly, and implied values ​​are generated using the mean / median of the historical data for that attribute, which are then overwritten and stored.

[0051] Step 105 Vector Clock Update and Asynchronous Synchronization After the central service completes the write operation, it updates the vector clock of the resource. The proxy service layer asynchronously synchronizes the same operation with the autonomous service. Upon receiving confirmation of successful execution from the central service, the proxy service layer updates its local vector clock.

[0052] Knowledge Request Process like Figure 3 As shown, when an edge node of an intelligent agent requests knowledge data from the central service, the proxy service layer implements adaptive multi-level scheduling based on a cost function according to the knowledge semantic descriptor. The specific steps are as follows: Step 201 Local Knowledge Base Query After receiving a knowledge retrieval task, the agent edge node first queries its local knowledge base. If a result is found and the vector clock verification passes (i.e., the local version is not strictly behind the central version), it returns the result directly. If no result is found or the version is outdated, the agent edge node constructs a query request and sends it to the agent service layer along with the KSD.

[0053] Step 202 Adaptive Scheduling Decision Based on Cost Function After receiving a request, the proxy service layer no longer queries in a fixed hierarchical order, but instead calculates the scheduling cost of each level and dynamically selects the optimal query level.

[0054] in, For the first The expected response latency cost of Tier 1 services (estimated using historical exponential moving averages). For from the first Level to acquire resources The estimated bandwidth consumption cost; For the first Level holding resources The cost of freshness; For the weighting coefficients, satisfying .

[0055]

[0056] in, , Resources for central services Vector clock, For nodes Locally held resources Vector clock, For the set minimum value, Current timestamp For nodes Last successful resource synchronization timestamp, Represents vector norm The optimal query level is selected as follows:

[0057] Among them, level 1 is the proxy service layer cache, level 2 is the autonomous service layer, and level 3 is the central service layer.

[0058] Step 203 Request Flight Deduplication After selecting the target level, query the flight dictionary. If the same request (with the combination of resource_uri and operation_semantics as the key) is being executed, wait for the existing request to complete and share the result; otherwise, create a new flight request.

[0059] Step 204 Adaptive Cache Verification For cached query results, perform double validity validation:

[0060] A cached entry is valid if and only if the adaptive TTL has not expired and the cached vector clock is not strictly behind the central service. The adaptive TTL is calculated as follows:

[0061] in, Based on the expiration time, This is the semantic entropy decay coefficient (the higher the semantic entropy, the shorter the TTL). This is the volatility decay coefficient (the more frequent the changes, the shorter the TTL). Representing knowledge resources The semantic entropy statistical feature value, For resources The total number of operations within a time window, when hour, .

[0062] Step 205 Semantic prefetching When requesting resources At that time, the proxy layer predictively prefetches relevant resources based on the semantic relationships in the knowledge graph:

[0063] Semantic similarity is calculated using cosine similarity.

[0064] in, Represents a local knowledge set. The preset similarity threshold is set to 0.75 in this embodiment. When the threshold is below 0.75, the prefetch recall rate improvement is limited, but bandwidth consumption increases significantly. When the threshold is above 0.75, the recall rate decreases significantly. The prefetched resource set is returned along with the main request result to reduce the number of subsequent requests.

[0065] Step 206 Asynchronous bidirectional backfilling After the central service or autonomous service returns valid data, the agent layer asynchronously populates the data back into the local knowledge base of the autonomous service and the agent edge nodes, while simultaneously updating the vector clocks at each level. Subsequent identical requests can then directly retrieve the data from the local knowledge base of the autonomous service or the agent edge nodes.

[0066] Step 207 Active cache invalidation Once the central service completes a knowledge update or deletion operation, it broadcasts a cache invalidation notification (carrying the changed resource_uri and the new vector clock) to all proxy layers. Upon receiving the notification, the proxy layer immediately deletes the entries matching the resource identifier from its local cache, ensuring cache consistency.

[0067] Knowledge proactive distribution function like Figure 4 As shown, the central service can proactively push knowledge data to groups of intelligent agent edge nodes or individual intelligent agent edge nodes based on different triggering conditions, through semantic entropy accumulation monitoring and node status awareness. The specific steps are as follows: (a) Differentiated synchronization of node activation perception Step 301 Activate state perception When an agent edge node starts up or resumes, it sends an activation registration request to the central service, carrying a KSD (containing resource_uri, consistency_stamp, semantic_vector, etc.).

[0068] Step 302 Synchronization strategy decision based on semantic consistency deviation After receiving the activation registration request, the central service calculates the semantic consistency deviation of the node:

[0069] in, and They are nodes and resources in the center service semantic vectors, The semantic similarity threshold is set to 0.85 in this embodiment. This means that two knowledge items are considered semantically equivalent when the cosine similarity of their semantic vectors is greater than 0.85. Represents a node The knowledge set held This represents the complete set of knowledge held by the central service. Represents cosine semantic similarity. .

[0070] Dynamically select synchronization strategy based on semantic consistency deviation:

[0071] in, To achieve full synchronization, all knowledge related to the edge nodes of the intelligent agent is extracted and distributed. For incremental synchronization, causal increments are extracted based on vector clocks and then distributed. For differential synchronization, only resources whose semantic entropy accumulation exceeds a threshold are synchronized, minimizing the amount of data transmitted. In this embodiment of the invention, the full synchronization threshold is... The value is 0.6, which is the incremental synchronization threshold. The value is 0.2, which means that when the semantic deviation exceeds 60%, a full reconstruction is triggered; when it is between 20% and 60%, incremental synchronization is triggered; and when it is below 20%, only differential repair is required.

[0072] Step 303 Incremental Extraction Based on Vector Clock For incremental synchronization, extract all resources in the central service whose vector clocks are not behind the known version of the target node:

[0073] The partial order relation of the vector clock is defined as follows:

[0074]

[0075] For differential synchronization, the semantic entropy accumulation exceeds a threshold. The resources are synchronized; in this embodiment, the value is taken as... That is, only when the historical cumulative semantic entropy of the resource exceeds Only then will it be included in the differential synchronization range:

[0076] in, This indicates the accumulation of its semantic entropy. Represents a node The incremental knowledge set, i.e., the central service decision node n A candidate set of knowledge resources that need updating but have not yet been synchronized.

[0077] Step 304 Activation Synchronization Completion Confirmation: After the agent's edge node completes knowledge merging, it sends a confirmation (carrying the updated vector clock) back to the central service, and the central service updates the knowledge version record of that node.

[0078] (ii) Change awareness triggered by semantic entropy accumulation Step 401: Monitoring Semantic Entropy Accumulation The central service maintains the cumulative semantic entropy value for each resource. Whenever a resource... When a change occurs, its semantic entropy is accumulated:

[0079] in, Representing resources At the last time of change The cumulative semantic entropy value; when the cumulative semantic entropy exceeds the distribution threshold... The active delivery is triggered at a specific time; in this embodiment, the value is taken as... That is, when the cumulative semantic entropy of resource r exceeds Active distribution is triggered at certain times.

[0080] Step 402: Priority sorting When multiple resources need to be distributed simultaneously, they are sorted in descending order of their accumulated semantic entropy values:

[0081] High semantic entropy resources are prioritized for distribution to ensure that high-impact changes reach each node first.

[0082] Step 403 Target Node Filtering and Differentiated Push The central service filters the list of affected active nodes based on the semantic vector and dependency graph of the resources, and selects a push strategy based on the semantic consistency deviation of each node (see step 302).

[0083] Step 404: High-priority forwarding at the proxy layer. When the agent layer recognizes that the adaptive_hint in KSD is BROADCAST, it prioritizes scheduling the distribution request, skips the ordinary cache queue, and directly pushes the data to the edge nodes of each target agent.

[0084] Step 405: Distribute and merge locally After receiving the data, the edge nodes of the intelligent agent perform a local knowledge base merging operation. Data fields are merged by key; knowledge entries are matched by resource_uri, updating the field if it exists and appending it if it does not; relationship edges are matched by head and tail identifiers, updating the field if it exists and appending it if it does not; after merging, the local vector clock is updated and persisted to the local knowledge base.

[0085] (III) Task-driven targeted knowledge distribution Step 501 Task Definition and Semantic Annotation The central service receives task instructions and generates a task distribution request. In addition to common fields, the KSD also carries the task_context (including task identifier, priority, and deadline). The request body describes the knowledge set conditions required for the target task.

[0086] Step 502 Target Agent Analysis The central service queries the node registry based on task conditions and filters a list of intelligent agent edge nodes that meet the requirements and are online.

[0087] Step 503 Knowledge Snapshot Assembly The central service extracts nodes and relation edges that meet the conditions from the full knowledge base and uses standardized knowledge data as the response body.

[0088] Step 504: High-priority forwarding at the proxy service layer. The proxy service layer pushes high-priority tasks directly, skipping the normal cache queue, based on the priority field of task_context in KSD.

[0089] Step 505: Distribute and merge locally After receiving the data, the edge nodes of the intelligent agent perform a local knowledge base merging operation. Data fields are merged by key; knowledge entries are matched by resource_uri, updating the field if it exists and appending it if it does not; relationship edges are matched by head and tail identifiers, updating the field if it exists and appending it if it does not; after merging, the local vector clock is updated and persisted to the local knowledge base.

[0090] This invention uses semantic entropy quantization to drive differentiated routing for knowledge reporting, cost-aware scheduling for knowledge requests, cumulative triggering of knowledge distribution, and three-level synchronization, and uses vector clocks to ensure causal consistency, thereby improving system processing efficiency, reducing latency, reducing bandwidth consumption, and ensuring data consistency.

[0091] In summary, the above are merely preferred embodiments of the present invention and are not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A multi-level knowledge management method based on semantic request-driven approach, characterized in that, include: Semantic entropy definition: For a multi-node knowledge graph system, the edge nodes of an agent maintain change markers and semantic vector indexes for the local knowledge base. For a single change operation of a knowledge resource, the semantic entropy of the knowledge change is defined to quantify the degree of semantic impact of a single change on the distributed system. Knowledge reporting: The edge nodes of the intelligent agent locally change knowledge data and construct semantic impact based on the semantic entropy of knowledge changes. The three dimensions of historical change frequency of knowledge resources, semantic entropy of current knowledge changes, and global dependence are integrated into a unified triggering quantitative indicator to replace the traditional reporting triggering logic that only relies on data volume or timestamp. When the semantic impact exceeds the adaptive reporting threshold, the agent's edge nodes are triggered to report the locally changed knowledge data to the central service. Knowledge Request: When an edge node of an agent requests knowledge data from the central service, the agent service layer implements an adaptive multi-level scheduling decision based on a cost function according to the knowledge semantic descriptor, and uses semantic entropy to calculate the adaptive time-to-live (TTL) to perform cache verification. Knowledge distribution: The central service maintains a semantic entropy accumulation value for each knowledge resource and only triggers active distribution when the accumulation value crosses the distribution threshold; by calculating the overall deviation between the knowledge held by the computing nodes and the central full knowledge in the semantic vector space, it dynamically selects full synchronization, incremental synchronization or differential synchronization.

2. The multi-level knowledge management method based on semantic request-driven according to claim 1, characterized in that, The semantic entropy of the knowledge change is calculated from the influence probability. In the process of constructing the influence probability, the dependency weights of nodes on resources in the knowledge graph topology and the real-time node activity status are integrated, so that the entropy value can dynamically reflect the semantic influence distribution of the change operation among active downstream nodes.

3. The multi-level knowledge management method based on semantic request-driven according to claim 2, characterized in that, For knowledge resources At any moment A change operation The semantic entropy of knowledge change is defined as: in, For change operation For the The probability of influence of each downstream consumption node This represents the total number of downstream consumption nodes.

4. The multi-level knowledge management method based on semantic request-driven according to claim 3, characterized in that, The probability of influence for: in, It is a very small number. For nodes Resources Dependency weights For a moment The set of nodes that are active, 1[ ] is an indicator function.

5. The multi-level knowledge management method based on semantic request-driven according to claim 4, characterized in that, The semantic influence for: in, Representing resources Normalized value of historical change frequency Representing resources The degree of dependence, This depends on the amplification factor.

6. The multi-level knowledge management method based on semantic request-driven approach according to claim 1, characterized in that, The proxy service layer implements adaptive multi-level scheduling decisions based on cost functions according to knowledge semantic descriptors, specifically as follows: Calculate the scheduling cost at each level. for: in, For the first The expected response delay cost of a service of the highest level. For from the first Level to acquire resources The estimated bandwidth consumption cost; For the first Level holding resources The cost of freshness; These are the weighting coefficients; The optimal query level is selected as follows: Among them, level 1 is the proxy service layer cache, level 2 is the autonomous service layer, and level 3 is the central service layer.

7. The multi-level knowledge management method based on semantic request-driven according to claim 4, characterized in that, The specific process of using semantic entropy to calculate adaptive TTL and perform cache verification is as follows: The adaptive TTL is calculated as follows: in, Based on the expiration time, This is the semantic entropy decay coefficient. This is the volatility attenuation coefficient. Representing knowledge resources The semantic entropy statistical feature value, For resources The total number of operations within a time window, when hour, , For resources The normalized value of the historical change frequency.

8. The multi-level knowledge management method based on semantic request-driven according to claim 1, characterized in that, When multiple resources need to be distributed simultaneously, they are sorted in descending order of their cumulative semantic entropy value, with resources of higher semantic entropy being distributed first.

9. The multi-level knowledge management method based on semantic request-driven according to claim 1, characterized in that, The degree of deviation between the knowledge held by computing nodes and the total knowledge of the center in the semantic vector space, i.e., semantic consistency deviation. for: in, and They are nodes and resources in the center service semantic vectors, For semantic similarity threshold, Represents a node The knowledge set held This represents the complete set of knowledge held by the central service. Indicates cosine semantic similarity; Dynamically select synchronization strategy based on semantic consistency deviation: in, To achieve full synchronization, all knowledge related to the edge nodes of the intelligent agent is extracted and distributed. For incremental synchronization, causal increments are extracted based on vector clocks and then distributed. For differential synchronization, only resources whose semantic entropy accumulation exceeds a threshold are synchronized, minimizing the amount of data transmitted. This indicates the full synchronization threshold. This indicates the incremental synchronization threshold.

10. The multi-level knowledge management method based on semantic request-driven according to claim 1, characterized in that, It also includes task-driven targeted knowledge delivery, where the knowledge semantic descriptor carries task identifiers, priorities, and deadlines.

Images