Knowledge full life circle management system and construction method

Abstract

The invention relates to the technical field of knowledge management systems, and discloses a knowledge full-life-cycle management system and a construction method, and the system comprises a technical platform and multi-modal perception layer, a calculation and arrangement engine layer, a unified API gateway layer, a multi-modal processing module layer, a knowledge graph layer, a cognitive reasoning layer and an application layer. Real-time collection and batch processing of multi-source heterogeneous data are achieved through a standardized SDK / API; spark / Flink is combined with Kubernetes to complete the ETL (Extract Transform Load) and resource scheduling of the multi-modal data; constructing an entity-relationship-attribute knowledge graph through cross-modal feature fusion; intelligent decision support is realized based on rule reasoning, graph calculation and GNN; the application layer provides intelligent questioning and answering, decision support and personalized recommendation services; the technical problems of knowledge islands, sharing barriers, knowledge statics and the like in the prior art are solved.

Description

Technical Field

[0001] This invention relates to the field of knowledge management system technology, specifically to a knowledge lifecycle management system and its construction method. Background Technology

[0002] As knowledge systems become more diversified and deeply integrated, the demand for knowledge management is increasingly prominent across various fields and industries. This trend requires entities not only to deepen their own knowledge reserves and improve their internal systems, but also to proactively integrate external knowledge and cognition to build a comprehensive knowledge perspective. However, existing technologies face the following core problems: Knowledge silos and barriers to sharing: Knowledge is scattered across departments, systems, platforms, and personal devices, forming information silos. A large amount of tacit knowledge (such as expert experience and decision-making logic) is difficult to explicitly recognize and retain, and employee turnover leads to the loss of knowledge assets. The causes include departmental barriers, system heterogeneity, and the lack of a unified knowledge management platform.

[0003] Static knowledge and superficial utilization: Unstructured data (PDF / Word / PPT / email / images / videos) cannot be automatically processed and analyzed by machines, heavily relying on human intervention. This is due to a lack of multimodal data fusion and processing capabilities.

[0004] Inefficient maintenance and biased understanding: Knowledge updates and maintenance rely on manual processes, and version inconsistencies lead to decreased accuracy; the system cannot simultaneously understand multimodal information such as text, charts, curves, and images. The causes include a lack of dynamic update mechanisms and cross-modal cognitive fusion capabilities.

[0005] The combined effect of these problems leads to an increased risk of knowledge asset loss, a decline in the accuracy of decision-making, and high knowledge management costs, which seriously restricts the process of intelligent transformation of enterprises. Summary of the Invention

[0006] The present invention aims to provide a knowledge lifecycle management system and its construction method to solve the technical problems of knowledge silos, sharing barriers and static knowledge in the existing technology.

[0007] To achieve the above objectives, the present invention adopts the following technical solution: a knowledge lifecycle management system, comprising: The technology platform and multimodal perception layer are used to provide reusable core technology capabilities and process multimodal data; The computation and orchestration engine layer is used for batch and real-time processing of multimodal data and resource scheduling is achieved through a container orchestration platform. The unified API gateway layer is used to provide REST / GraphQL interfaces to the outside world and integrate authentication, authorization, rate limiting, monitoring and log management functions; The multimodal processing module layer is used for feature extraction and cross-modal semantic fusion of text, image, audio, and video data; The knowledge graph layer is used to construct a knowledge network in the form of entity-relationship-attribute after the transformation of structured, semi-structured and multimodal data; The cognitive reasoning layer is used to perform rule-based reasoning, graph computation, graph neural network reasoning, and intelligent question answering based on knowledge graphs. The application layer provides intelligent application services such as intelligent question answering, decision support, search, and recommendation.

[0008] The principle and advantages of this solution are as follows: In practical applications, it connects to multiple heterogeneous data sources through a standardized SDK / API, and combines real-time incremental data collection via CDC with distributed task scheduling via Airflow / DolphinScheduler to achieve cross-department / system data synchronization and unified storage. This breaks down departmental barriers and knowledge silos caused by system heterogeneity, solving the problem of tacit knowledge accumulation; real-time collection improves data timeliness, and categorized storage reduces retrieval latency, breaking down barriers to knowledge sharing.

[0009] By leveraging batch and stream collaborative processing, we improve knowledge processing efficiency and address the problem of static utilization of unstructured data; containerized resource scheduling optimizes resource utilization, resolving resource waste caused by static processing. A unified service entry point and security protection layer address access control and security issues in cross-system knowledge sharing, reducing sharing barriers. Cross-modal semantic fusion addresses cognitive biases, enhances the machine's automated processing capabilities for unstructured data, and resolves the problem of static knowledge utilization. Structured knowledge networks address knowledge silos and static storage issues; version management reduces maintenance costs and resolves version chaos caused by manual knowledge updates. Intelligent reasoning and decision support address declining decision-making accuracy; application-layer services realize the value of knowledge and address high knowledge management costs.

[0010] Preferably, as an improvement, the technology platform and multimodal perception layer include: The multi-source data acquisition module supports integration with databases, file systems, APIs, and multimodal data sources via standardized SDKs / APIs. The Modify Data Capture (CDC) module is used to achieve real-time incremental data acquisition; The distributed task scheduling module, based on DolphinScheduler or Airflow, implements scheduled and real-time task orchestration. The system includes a categorized storage module, comprising HDFS / S3 for raw data, a graph database for knowledge graphs, a vector database for multimodal feature vectors, and Elasticsearch for full-text search.

[0011] The benefits of this improvement are as follows: It enables cross-source data integration through standardized SDKs / APIs, and combines real-time data acquisition from the CDC with Airflow task orchestration to build an integrated "acquisition-storage-processing" workflow. The categorized storage module uses HDFS / S3 to store raw data, graph databases to store knowledge graphs, and vector databases to store feature vectors, forming a multimodal data lake. This breaks down departmental barriers and system heterogeneity limitations, achieving unified knowledge management; real-time incremental acquisition improves data timeliness; and categorized storage reduces data retrieval latency, solving the problems of knowledge silos and static storage.

[0012] Preferably, as an improvement, the computation and orchestration engine layer includes: Spark batch processing engine is used to execute multimodal ETL tasks; The Flink real-time processing engine is used to perform multimodal real-time parsing tasks; Kubernetes is a container orchestration platform used to implement resource scheduling and elastic scaling of computing tasks.

[0013] The benefits of this improvement are: the Spark batch processing engine executes multimodal ETL tasks, the Flink real-time processing engine completes streaming data parsing, and Kubernetes dynamically schedules container resources based on task load. Batch and streaming collaborative processing improves knowledge processing efficiency; containerized resource scheduling improves resource utilization and solves the problems of static processing and resource waste.

[0014] Preferably, as an improvement, the multimodal processing module layer includes: A text analysis engine used to perform NLP tasks such as NER, RE, and EL; The image parsing engine extracts image information and associates it with graph entities based on CNN, YOLO, and OCR. The audio and video processing engine, based on ASR and voiceprint recognition, enables speech-to-text conversion and speaker identification. The multimodal feature fusion module is used to map text, image, and audio features to a unified semantic space and enhance the entity dimension of the knowledge graph.

[0015] The beneficial effects of this improvement are as follows: the text analysis engine extracts entity relationships through NER / RE / EL; the image parsing engine uses CNN / YOLO to identify defect features and associate them with graph entities; the audio and video processing engine uses ASR to convert text and voiceprint recognition to identify the speaker; and the multimodal feature fusion module uses cross-modal Transformer to map features to a unified semantic space. Entity recognition accuracy is high, and problem diagnosis efficiency is improved; voice emotion analysis and speaker recognition enhance customer feedback and provide a more comprehensive understanding, addressing the problem of partial cognition.

[0016] Preferably, as an improvement, the knowledge graph layer includes: The graph schema layer is used to construct entities, relationships, and attribute types through ontology definitions, and provides a UI for domain experts to design and review. The graph construction module is used to realize entity recognition, relation extraction, entity alignment and knowledge fusion based on Spark NLP / Flink-AI; The graph storage and management module is used to import knowledge into the graph database and supports version management, snapshots, and rollback functions.

[0017] The benefits of this improvement are as follows: the graph schema layer constructs an entity-relationship-attribute framework through ontology definition and provides a UI for expert review; the construction module completes the entire knowledge extraction process based on Spark NLP / Flink-AI; and the storage module supports version management, snapshots, and rollback. Standardized knowledge construction improves consistency; entity alignment technology resolves multi-source conflicts and improves the accuracy of knowledge updates; and version management reduces the risk of version chaos and solves the problem of inefficient maintenance.

[0018] Preferably, as an improvement, the cognitive reasoning layer includes: The rule reasoning module performs business rule matching and decision-making based on the Drools engine; The graph computing inference module performs path query and community discovery algorithms based on Neo4j GDS or GraphX. The graph neural network inference module implements link prediction and node classification based on PyTorch Geometric or DGL. An intelligent question-answering engine is used to convert natural language questions into graph query statements and return accurate answers.

[0019] The beneficial effects of this improvement are: the rule reasoning module matches business rules using Drools; the graph computing module performs path querying and community discovery; the GNN reasoning module implements link prediction and node classification; and the intelligent question answering engine converts natural language into Cypher queries. Rule matching response time is <100ms; graph algorithms mine deep relationships to improve decision-making depth; and GNN supplements missing relationships to improve entity attributes, solving the problem of shallow cognition.

[0020] Preferably, as an improvement, the application layer includes: Intelligent question-answering systems are used to provide interactive natural language query services based on knowledge graphs; Decision support systems are used to generate business decision recommendations based on rules and graph computation results; Intelligent search and recommendation systems are used to achieve personalized search and content recommendation based on graph structure and multimodal features.

[0021] The beneficial effects of this improvement are: the intelligent question-answering system achieves semantic retrieval based on knowledge graphs; the decision support system integrates rules and graph computation results; and the intelligent recommendation system associates user behavior with graph entities. Question-answering response time is less than 100ms, improving decision accuracy; personalized recommendations enhance user satisfaction and solve the challenges of application implementation.

[0022] Preferably, as an improvement, a dynamic priority calculation formula is introduced into Airflow task scheduling, which combines three dimensions: task urgency, data volume, and resource requirements, weighted together. ; in, This represents the dynamic priority of the task; the higher the value, the higher the priority. This is the urgency level coefficient. Predefined by business rules; This represents the data volume, in GB. The logarithm is used to prevent the influence of dimensions. Real-time statistics are collected via the data acquisition module; For resource requirements, the number of cores per unit is used, and the square root is taken to smooth out extreme values. Estimate resource requirements based on task resource requirements documents; The weighting coefficients are dynamic and are adjusted by monitoring node CPU utilization and memory pressure in real time. The weighting coefficients are dynamically calculated by collecting node load data every 5 minutes through the Prometheus monitoring system.

[0023] The benefits of this improvement are: dynamic priority is calculated by weighting three dimensions—urgency, data volume, and resource requirements—with the weighting coefficients dynamically adjusted according to node load. This shortens response time for urgent tasks, improves resource utilization, and solves the problem of inefficient task scheduling.

[0024] A method for constructing a knowledge lifecycle management system, comprising: Build a technology platform and multimodal perception layer: Connect to databases, API interfaces, file systems, message queues and SaaS applications through a standardized SDK / API connector framework, support real-time incremental data collection and scheduled batch data collection for CDC, and perform task orchestration and dependency management based on Airflow to achieve efficient synchronization of multi-source heterogeneous data; Establish a unified data storage system: use HDFS or object storage to store the original data, select Neo4j, NebulaGraph or JanusGraph to store the knowledge graph, deploy Milvus or Chroma to store multimodal feature vectors, and combine Elasticsearch to realize full-text and similarity retrieval, forming a multimodal and multi-level storage architecture. Build a computing and orchestration engine: Implement large-scale multimodal data ETL based on Spark, complete real-time data stream processing and knowledge updates through Flink combined with Kafka / Pulsar, and implement containerized resource scheduling and elastic scaling based on Kubernetes; Deploy a unified API gateway: Provide REST / GraphQL dual-mode interfaces to the outside world, integrate OAuth2.0 authentication, RBAC authorization, token bucket rate limiting, Prometheus monitoring and ELK log management, forming a unified service entry point and security protection layer; Achieve multimodal cognitive fusion: Integrate Spark NLP / Flink-AI to complete entity recognition and relation extraction, extract image information and associate it with graph entities based on CNN / YOLO and OCR, process speech data through ASR and voiceprint recognition, and use contrastive learning or cross-modal Transformer to map text, image and audio features to a unified semantic space to achieve cross-modal association fusion; Construct a knowledge graph layer: Define an ontology framework and provide a visual UI for domain experts to design and review. Rely on SparkNLP / Flink-AI to realize the entire process of preprocessing, NER, relation extraction, attribute extraction, and entity linking. Combine MLflow / Kubeflow to manage model versions and deployment. Resolve multi-source entity conflicts through entity alignment, realize cross-data source knowledge fusion, and import structured knowledge into a graph database. Support version management, snapshots, and rollback. Build a cognitive reasoning layer: integrate the Drools engine to perform business rule reasoning, execute graph algorithms based on Neo4j GDS or SparkGraphX ​​to mine deep associations, use PyTorch Geometric / DGL to implement graph neural network link prediction and node classification, and transform natural language questions into Cypher queries through BERT fine-tuning to achieve intelligent question answering and multi-turn dialogue; Integrated application layer functionality: It calls underlying services through a unified API gateway, combines Elasticsearch full-text search with knowledge graph entity linking to achieve semantic-level search, generates business decision suggestions based on rule reasoning and graph computation results, and uses user behavior data and knowledge graph entity association to achieve personalized content recommendation, completing the entire lifecycle closed loop from data collection to intelligent application.

[0025] The benefits of this improvement are: a closed-loop design throughout the entire process enables seamless integration from data collection to application; multimodal feature fusion enhances the dimensionality of the data graph; and a unified API gateway integrates security control functions. Knowledge management efficiency and comprehensiveness are improved; service entry security is enhanced, solving the challenges of end-to-end management. Attached Figure Description

[0026] Figure 1 This is a schematic diagram of the structure of an embodiment of the present invention. Detailed Implementation

[0027] The following detailed description illustrates the specific implementation method: Example The basics are as follows: Figure 1 As shown, a knowledge lifecycle management system and its construction method are disclosed, specifically including a knowledge lifecycle management system and its construction method.

[0028] A knowledge lifecycle management system, comprising: The technology platform and multimodal perception layer are not only the underlying technological foundation of the system, providing reusable core technical capabilities to support upper-layer modules; they can also break through the limitations of text and process multimodal data, including images, audio and video. Through feature fusion, they enhance the entity dimension of the knowledge graph, achieving the dual function of "technical support + perception upgrade".

[0029] At the data acquisition level, the technology platform and multimodal perception layer connect to all types of data sources through standardized software development kits (SDKs) and application programming interfaces (APIs). This covers traditional data sources such as databases (e.g., Java Database Connectivity (JDBC) / Open Database Connectivity (ODBC)), APIs, and file systems (e.g., Amazon Simple Storage Service (S3) / Hadoop Distributed File System (HDFS)), as well as multimodal data source access. Combining real-time incremental acquisition and scheduled batch acquisition modes with Change Data Capture (CDC), and relying on the distributed workflow scheduling system DolphinScheduler / open-source workflow scheduling platform Airflow for task orchestration, efficient acquisition of multi-source data is ensured.

[0030] In Airflow task scheduling, a dynamic priority calculation formula is introduced, which combines three dimensions: task urgency, data volume, and resource requirements. ; in, This represents the dynamic priority of the task; the higher the value, the higher the priority. This is the urgency level coefficient. Predefined by business rules; The data volume (in GB) is represented by a logarithm to prevent the influence of dimensions. Real-time statistics are collected via the data acquisition module; For resource requirements (number of cores per unit), take the square root to smooth out extreme values. Estimate resource requirements based on task resource requirements documents; The weighting coefficients are dynamic and are adjusted by monitoring node CPU utilization and memory pressure in real time. The weighting coefficients are dynamically calculated by collecting node load data every 5 minutes through the Prometheus monitoring system.

[0031] For example, an e-commerce company's customer feedback system processes over 20,000 product reviews and complaints daily, requiring dynamic priority scheduling and retrieval optimization to achieve efficient processing and rapid querying.

[0032] The following settings can then be made: (1) Urgent complaint task: urgency level coefficient : 0.9, requiring a response within 24 hours; data volume Average 500 characters / entry (taking...) Resource requirements 8-core CPU / stick (take) ); Dynamic weighting coefficients The values ​​are 0.9, 0.1, and 0.2 respectively. Under high load... Automatically upgraded to 0.9.

[0033] (2) General Comment Task: Urgency level coefficient : 0.5, requiring a response within 72 hours; data volume Average 200 characters / entry (taking...) Resource requirements 4-core CPU / stick (take) ); Dynamic weighting coefficients The values ​​are 0.9, 0.1, and 0.2 respectively. Under high load... Automatically upgraded to 0.9.

[0034] Based on this calculation, the dynamic priority of emergency complaint tasks is 1.65, while that of ordinary comment tasks is 1.16. The priority of complaint tasks is 42% higher, ensuring that emergency complaints are given priority in the processing queue, and the average waiting time can be reduced by 60%.

[0035] In traditional systems, urgent complaints (such as product quality defects) and ordinary comments (such as feature suggestions) are stored separately, lacking a cross-departmental collaborative processing mechanism. Through dynamic priority scheduling, urgent complaints can automatically trigger cross-departmental knowledge linkage, avoiding response delays caused by departmental silos. Customer feedback includes multimodal data such as text comments, product images, and voice complaints. The system uses feature fusion technology to associate product defect features in images (such as crack detection), emotional features in voice (such as angry tone), and text comments, constructing a three-dimensional knowledge graph containing "entity-relationship-attribute". To address the issue of inconsistent knowledge versions, the system employs a dynamic weight adjustment mechanism. When high load is detected (such as a surge in complaints during promotional seasons), the urgency level coefficient is automatically increased. Ensure that high-priority tasks occupy resources first; increase the weight of data volume dimension during low load. Optimize batch processing efficiency.

[0036] The unified data storage of the technology platform and the multimodal perception layer adopts a categorized storage strategy: HDFS / S3 stores raw data and multimodal raw files, high-performance graph databases such as Neo4j / Nebula Graph (a domestic open-source graph database) store knowledge graph data, vector similarity search engine Milvus / open-source vector database Chroma stores multimodal feature vectors, and the distributed search and analysis engine Elasticsearch undertakes full-text search functions, forming a multi-layered, multimodal storage system.

[0037] The computing and orchestration engine is based on the Spark batch processing engine (including multimodal ETL) and Flink real-time processing engine (including multimodal real-time parsing), combined with the Kubernetes containerized resource scheduling platform to ensure the efficient operation and resource optimization of multimodal computing tasks.

[0038] The unified API gateway provides external API interfaces that represent state transitions using REST / API query language (GraphQL), covering traditional business and multimodal service calls. It also integrates authentication, authorization, rate limiting, monitoring, and log management functions, forming a unified service entry point and security protection layer.

[0039] The multimodal processing module, as the core of cognitive fusion, includes a text analysis engine, an image parsing engine, and an audio / video processing engine. Ultimately, through multimodal feature fusion, it maps the modal features of text, images, audio, and other modalities to the same semantic space, achieving cross-modal association fusion, further enriching the entity dimension of the knowledge graph, and enhancing the system's perception and cognitive capabilities.

[0040] The text analysis engine supports the knowledge graph's Natural Language Processing (NLP) capabilities, providing fundamental text support. The image analysis engine extracts image information and associates it with knowledge graph entities based on Convolutional Neural Networks (CNN), the YOLO object detection algorithm, and Optical Character Recognition (OCR) technology. The audio / video processing engine converts meeting recordings into text using Automatic Speech Recognition (ASR) and sends it for analysis, combining this with voiceprint recognition to determine the speaker's identity.

[0041] The knowledge graph layer, as the core knowledge foundation for intelligent reasoning in the system, receives the processing results from the technology platform and the multimodal perception layer. It transforms unstructured and semi-structured data (including data after multimodal transformation) into structured knowledge of "entity-relationship-attribute", thereby realizing the structured reconstruction of knowledge and supporting intelligent applications.

[0042] Its core consists of three main modules: The graph schema layer constructs a knowledge framework through ontology definition, clearly defining entity, relation, and attribute types, and providing a user interface (UI) for domain experts to participate in design and review, ensuring the standardization and domain adaptability of the knowledge system; The graph construction module relies on Spark NLP / Flink-AI to implement an NLP pipeline, covering the entire process of text preprocessing, entity recognition (NER), relation extraction (RE), attribute extraction, and entity linking (EL). It combines the machine learning lifecycle management platform MLflow / open-source machine learning platform Kubeflow to manage model training and deployment, simultaneously resolving entity conflicts in multi-source data (including multimodal data), and achieving cross-data source knowledge fusion through entity alignment technology; The graph storage and management system imports structured knowledge into graph databases (such as Neo4j / JanusGraph), supports version management, snapshots, and rollback functions, ensuring the stability and traceability of the knowledge graph, and ultimately forming an intelligent knowledge foundation covering the entire knowledge lifecycle.

[0043] The cognitive reasoning layer, as the "brain" of the system, generates decision-level intelligent results based on the fusion of structured knowledge from the knowledge graph and information from the multimodal perception layer, through core capabilities such as rule-based reasoning, graph computation, and intelligent question answering, supporting the implementation of intelligent applications.

[0044] Its core is an intelligent reasoning engine that integrates multi-dimensional reasoning capabilities: the rule reasoning module relies on the Drools rule engine to process explicit business logic, such as "when there are more than 10 customer complaints and the product version is A, a quality warning is triggered", to achieve rapid rule matching and decision-making; The graph computation reasoning module uses the Neo4j GDS graph data science library or the Spark graph computation framework (GraphX) to perform graph algorithms such as path querying and community discovery to uncover deep relationships between entities. The graph neural network (GNN) reasoning module uses the PyTorch Geometric / Depth Graph Learning (DGL) framework to perform machine learning reasoning such as link prediction (i.e., filling in missing relationships) and node classification through graph neural networks to supplement missing relationships in the knowledge graph or improve entity attributes. The intelligent question-answering engine transforms natural language questions into graph query statements such as Cypher through semantic parsing, directly retrieves and returns accurate answers from the knowledge graph, and ultimately forms an intelligent reasoning closed loop of "rule-driven + algorithm mining + semantic interaction", which promotes the system's upgrade from knowledge storage to intelligent decision-making.

[0045] The application layer, relying on the capabilities of the aforementioned layers, provides specific applications such as intelligent question answering, decision support systems, and intelligent search and recommendation, forming a complete closed loop from data collection to intelligent application, supporting complex data analysis and decision-making needs. The entire system, through a multi-layered architecture design, achieves full-process coverage from technical support, perception upgrades, knowledge reconstruction, reasoning and decision-making to application implementation, constructing a complete knowledge lifecycle management system.

[0046] A method for constructing a knowledge lifecycle management system includes: Step 1: Construct a technology platform and multimodal awareness layer. Through a standardized SDK / API connector framework, connect to databases (JDBC / ODBC), API interfaces (RESTful), file systems (S3 / HDFS), message queues (Kafka / Pulsar), and SaaS applications to achieve access to all types of data sources. Support both real-time incremental data collection (such as database log capture) and scheduled batch collection modes. Use Airflow for task orchestration and dependency management to ensure efficient synchronization of multi-source data.

[0047] Step 2: Establish a unified data storage system, using HDFS or object storage (such as S3 / OSS) to store raw data, supporting columnar storage formats such as Parquet / ORC / AVRO, and optimizing query performance. Select Neo4j, Nebula Graph, or JanusGraph to store knowledge graph data, supporting the Cypher query language. Deploy Milvus or Chroma to store multimodal feature vectors, and combine with Elasticsearch to implement full-text search and similarity search functions.

[0048] Step 3: Build a computing and orchestration engine, implementing large-scale data ETL and multimodal data processing based on Spark, supporting custom pipeline development. Real-time data processing and knowledge updates are achieved through Flink combined with Kafka / Pulsar, supporting low-latency data stream processing. Kubernetes containerized scheduling enables elastic scaling of computing resources and task isolation.

[0049] Step 4: Deploy a unified API gateway to provide REST / GraphQL dual-mode interfaces, integrate OAuth2.0 authentication, RBAC authorization, token bucket rate limiting, Prometheus monitoring and ELK log management functions, forming a unified service entry point.

[0050] Step 5: Achieve multimodal cognitive fusion by integrating Spark NLP / Flink-AI to realize NLP capabilities such as entity recognition and relation extraction. Extract image information based on CNN / YOLO object detection and OCR technologies, and associate them with geographic entities. Convert speech to text using ASR, and combine this with speaker recognition to identify the speaker; video processing supports keyframe extraction and action recognition. Employ contrastive learning or a cross-modal Transformer to map text, image, and audio features to a unified semantic space, achieving cross-modal association and fusion.

[0051] Step 6: Construct the knowledge graph layer, define the ontology framework (entity / relationship / attribute type), and provide a visual UI for domain experts to design and review, ensuring the standardization and domain adaptability of the knowledge system. Integrate Spark NLP / Flink-AI to achieve the entire process of preprocessing → NER → relation extraction → attribute extraction → entity linking, and combine MLflow / Kubeflow to manage model versions and deployment; resolve entity conflicts in multi-source data through entity alignment technology to achieve cross-data source knowledge fusion. Import structured knowledge into a graph database, supporting version management, snapshots, and rollback functions to ensure the stability and traceability of the knowledge graph.

[0052] Step 7: Build a cognitive reasoning layer, integrating the Drools engine to handle explicit business logic (e.g., "more than 10 customer complaints and product version A → trigger quality warning"); perform graph algorithms such as path querying and community discovery based on Neo4j GDS or Spark GraphX; and implement GNN inference such as link prediction and node classification using the PyTorch Geometric / DGL framework. Transform natural language questions into Cypher graph query statements through semantic parsing (e.g., BERT fine-tuning), directly retrieving accurate answers from the knowledge graph, supporting multi-turn dialogue and intent recognition.

[0053] Step 8: Integrate application layer functionality. The application layer interacts with the system through a unified API gateway, calling services from the technology platform, knowledge graph layer, and cognitive reasoning layer. It combines Elasticsearch full-text search with knowledge graph entity linking to achieve semantic-level search. Business decision suggestions are generated based on rule-based reasoning and graph computation results. Personalized content recommendations are achieved by linking user behavior data with knowledge graph entities.

[0054] As attached Figure 1 As shown, the flow from bottom to top is unidirectional: the arrows are uniformly directed from bottom to top, reflecting the complete closed loop of "technical support → perception fusion → knowledge construction → reasoning and decision-making → application implementation".

[0055] The technology platform provides the data and computing power foundation for the multimodal cognitive fusion layer through multi-source data acquisition and unified storage; the multimodal cognitive fusion layer delivers structured knowledge raw materials to the knowledge graph layer through text / image / audio processing and feature fusion; the knowledge graph layer provides the knowledge foundation of "entity-relationship-attribute" for the cognitive reasoning layer through pattern definition, construction, and storage; and the cognitive reasoning layer ultimately drives intelligent decision-making in the application layer through rule-based reasoning, graph computing, and GNN.

[0056] Technology Platform → Multimodal Cognitive Fusion Layer: The technology platform's "unified API gateway" and "computation orchestration engine" directly support real-time acquisition of multimodal data (such as JDBC / ODBC protocols), distributed processing (Spark / Flink), and resource scheduling (Kubernetes), ensuring efficient access and preliminary processing of multimodal data (images, audio, and video).

[0057] Multimodal cognitive fusion layer → Knowledge graph layer: Multimodal processing modules (such as CNN image parsing and ASR speech-to-text) transform unstructured data into structured features, which are then mapped to a unified semantic space through feature fusion and finally input into the knowledge graph layer for entity relationship modeling.

[0058] For example, the emotional characteristics of voice in customer complaints are linked with the text description of defects to form a knowledge chain of "emotion-defect-solution", and the direction of the arrows reflects the dimensional upgrading process of "perception → cognition".

[0059] Knowledge Graph Layer → Cognitive Reasoning Layer: The knowledge graph layer defines the ontology framework (such as entity types for products, users, and defects) through the "graph pattern layer," generates structured knowledge through the "graph building module" (NER / RE / EL pipeline), and stores it in a graph database (Neo4j) or a vector database (Milvus) for the inference engine to access.

[0060] Cognitive reasoning layer → Application layer: The cognitive reasoning layer directly supports the application layer's intelligent question answering (Cypher query), decision support (quality warning), and recommendation system (user profile matching) through rule-based reasoning (such as the Drools engine), graph computing (community discovery), and GNN link prediction capabilities.

[0061] The above descriptions are merely embodiments of the present invention, and common knowledge such as specific technical solutions and / or characteristics are not described in detail here. It should be noted that those skilled in the art can make various modifications and improvements without departing from the technical solutions of the present invention, and these should also be considered within the scope of protection of the present invention. These modifications and improvements will not affect the effectiveness of the implementation of the present invention or the practicality of the patent. The scope of protection claimed in this application should be determined by the content of its claims, and the specific embodiments described in the specification can be used to interpret the content of the claims.

Claims

1. A knowledge lifecycle management system, characterized in that, include: The technology platform and multimodal perception layer are used to provide reusable core technology capabilities and process multimodal data; The computation and orchestration engine layer is used for batch and real-time processing of multimodal data and resource scheduling is achieved through a container orchestration platform. The unified API gateway layer is used to provide REST / GraphQL interfaces to the outside world and integrate authentication, authorization, rate limiting, monitoring and log management functions; The multimodal processing module layer is used for feature extraction and cross-modal semantic fusion of text, image, audio, and video data; The knowledge graph layer is used to construct a knowledge network in the form of entity-relationship-attribute after the transformation of structured, semi-structured and multimodal data; The cognitive reasoning layer is used to perform rule-based reasoning, graph computation, graph neural network reasoning, and intelligent question answering based on knowledge graphs. The application layer provides intelligent application services such as intelligent question answering, decision support, search, and recommendation.

2. The knowledge lifecycle management system according to claim 1, characterized in that, The technology platform and multimodal perception layer include: The multi-source data acquisition module supports integration with databases, file systems, APIs, and multimodal data sources via standardized SDKs / APIs. The Modify Data Capture (CDC) module is used to achieve real-time incremental data acquisition; The distributed task scheduling module, based on DolphinScheduler or Airflow, implements scheduled and real-time task orchestration. The system includes a categorized storage module, comprising HDFS / S3 for raw data, a graph database for knowledge graphs, a vector database for multimodal feature vectors, and Elasticsearch for full-text search.

3. A knowledge lifecycle management system according to claim 2, characterized in that, The computation and orchestration engine layer includes: Spark batch processing engine is used to execute multimodal ETL tasks; The Flink real-time processing engine is used to perform multimodal real-time parsing tasks; Kubernetes is a container orchestration platform used to implement resource scheduling and elastic scaling of computing tasks.

4. A knowledge lifecycle management system according to claim 3, characterized in that, The multimodal processing module layer includes: A text analysis engine used to perform NLP tasks such as NER, RE, and EL; The image parsing engine extracts image information and associates it with graph entities based on CNN, YOLO, and OCR. The audio and video processing engine, based on ASR and voiceprint recognition, enables speech-to-text conversion and speaker identification. The multimodal feature fusion module is used to map text, image, and audio features to a unified semantic space and enhance the entity dimension of the knowledge graph.

5. A knowledge lifecycle management system according to claim 4, characterized in that, The knowledge graph layer includes: The graph schema layer is used to construct entities, relationships, and attribute types through ontology definitions, and provides a UI for domain experts to design and review. The graph construction module is used to realize entity recognition, relation extraction, entity alignment and knowledge fusion based on Spark NLP / Flink-AI; The graph storage and management module is used to import knowledge into the graph database and supports version management, snapshots, and rollback functions.

6. A knowledge lifecycle management system according to claim 5, characterized in that, The cognitive reasoning layer includes: The rule reasoning module performs business rule matching and decision-making based on the Drools engine; The graph computing inference module performs path query and community discovery algorithms based on Neo4j GDS or GraphX. The graph neural network inference module implements link prediction and node classification based on PyTorch Geometric or DGL. An intelligent question-answering engine is used to convert natural language questions into graph query statements and return accurate answers.

7. A knowledge lifecycle management system according to claim 6, characterized in that, The application layer includes: Intelligent question-answering systems are used to provide interactive natural language query services based on knowledge graphs; Decision support systems are used to generate business decision recommendations based on rules and graph computation results; Intelligent search and recommendation systems are used to achieve personalized search and content recommendation based on graph structure and multimodal features.

8. A knowledge lifecycle management system according to claim 7, characterized in that: In Airflow task scheduling, a dynamic priority calculation formula is introduced, which combines three dimensions: task urgency, data volume, and resource requirements. ； in, This represents the dynamic priority of the task; the higher the value, the higher the priority. This is the urgency level coefficient. Predefined by business rules; This represents the data volume, in GB. The logarithm is used to prevent the influence of dimensions. Real-time statistics are collected via the data acquisition module; For resource requirements, the number of cores per unit is used, and the square root is taken to smooth out extreme values. Estimate resource requirements based on task resource requirements documents; The weighting coefficients are dynamic and are adjusted by monitoring node CPU utilization and memory pressure in real time. The weighting coefficients are dynamically calculated by collecting node load data every 5 minutes through the Prometheus monitoring system.

9. A method for constructing a knowledge lifecycle management system, used to construct a knowledge lifecycle management system as described in any one of claims 1-8, characterized in that, include: Build a technology platform and multimodal perception layer: Connect to databases, API interfaces, file systems, message queues and SaaS applications through a standardized SDK / API connector framework, support real-time incremental data collection and scheduled batch data collection for CDC, and perform task orchestration and dependency management based on Airflow to achieve efficient synchronization of multi-source heterogeneous data; Establish a unified data storage system: use HDFS or object storage to store the original data, select Neo4j, NebulaGraph or JanusGraph to store the knowledge graph, deploy Milvus or Chroma to store multimodal feature vectors, and combine Elasticsearch to realize full-text and similarity retrieval, forming a multimodal and multi-level storage architecture. Build a computing and orchestration engine: Implement large-scale multimodal data ETL based on Spark, complete real-time data stream processing and knowledge updates through Flink combined with Kafka / Pulsar, and implement containerized resource scheduling and elastic scaling based on Kubernetes; Deploy a unified API gateway: Provide REST / GraphQL dual-mode interfaces to the outside world, integrate OAuth2.0 authentication, RBAC authorization, token bucket rate limiting, Prometheus monitoring and ELK log management, forming a unified service entry point and security protection layer; Achieve multimodal cognitive fusion: Integrate Spark NLP / Flink-AI to complete entity recognition and relation extraction, extract image information and associate it with graph entities based on CNN / YOLO and OCR, process speech data through ASR and voiceprint recognition, and use contrastive learning or cross-modal Transformer to map text, image and audio features to a unified semantic space to achieve cross-modal association fusion; Construct a knowledge graph layer: Define an ontology framework and provide a visual UI for domain experts to design and review. Rely on SparkNLP / Flink-AI to realize the entire process of preprocessing, NER, relation extraction, attribute extraction, and entity linking. Combine MLflow / Kubeflow to manage model versions and deployment. Resolve multi-source entity conflicts through entity alignment, realize cross-data source knowledge fusion, and import structured knowledge into a graph database. Support version management, snapshots, and rollback. Build a cognitive reasoning layer: integrate the Drools engine to perform business rule reasoning, execute graph algorithms based on Neo4j GDS or Spark GraphX ​​to mine deep associations, use PyTorch Geometric / DGL to implement graph neural network link prediction and node classification, and transform natural language questions into Cypher queries through BERT fine-tuning to achieve intelligent question answering and multi-turn dialogue; Integrated application layer functionality: It calls underlying services through a unified API gateway, combines Elasticsearch full-text search with knowledge graph entity linking to achieve semantic-level search, generates business decision suggestions based on rule reasoning and graph computation results, and uses user behavior data and knowledge graph entity association to achieve personalized content recommendation, completing the entire lifecycle closed loop from data collection to intelligent application.

Images