Method and system for automatically constructing knowledge base in IT operation and maintenance field

Abstract

The invention belongs to the technical field of artificial intelligence, and discloses a method and system for automatically constructing a knowledge base in the IT operation and maintenance field, and the method comprises the steps: constructing an operation and maintenance knowledge model based on an RDF model; based on the constructed operation and maintenance knowledge model, performing analysis and knowledge extraction on original document data by using a large language model in combination with cue words; knowledge fusion is carried out on data of knowledge extraction quality, knowledge repetition of different data sources and association problems among knowledge; performing persistent storage and vector representation storage on a knowledge fusion result; and for the data stored in the vector representation, a graph path and vectorization mixed retrieval method is utilized, the semantic relation and the weight path relation are synthesized, and information required in operation and maintenance knowledge retrieval is obtained. The method can solve the problems existing in a traditional operation and maintenance knowledge base, the operation and maintenance efficiency and quality are improved, and powerful support is provided for intelligent operation and maintenance.

Description

Technical Field

[0001] The present invention belongs to the field of artificial intelligence technology, and in particular relates to a method and system for automatically constructing a knowledge base in the field of IT operation and maintenance. Background Art

[0002] With the rapid development of information technology and the accelerating pace of enterprise digital transformation, enterprise IT development is advancing at an unprecedented pace. The construction of IT infrastructure and diversified information systems is experiencing rapid growth. The diversification of IT infrastructure equipment and the continued expansion and complexity of business systems are creating unprecedented challenges for operations and maintenance.

[0003] Current IT operations and maintenance generally face the following challenges: increased system complexity and maintenance difficulty; enterprise IT construction involves diverse equipment and systems, significantly increasing the complexity of operations and maintenance; the mobility of technical personnel further exacerbates the difficulties of operations and maintenance, posing severe challenges to business continuity and stability.

[0004] Limited technical response capabilities. Traditional operations and maintenance models rely too heavily on individual skills and experience, lacking a systematic, standardized knowledge management system. This makes it difficult to effectively cope with complex and volatile operations and maintenance environments. Inadequate fault location and resolution times directly impact business stability and availability.

[0005] Inadequate data analysis capabilities: Data analysis relies on manual labor, resulting in low processing efficiency and limited depth by analysts' knowledge and experience. Comprehensive analysis tasks such as multi-dimensional analysis, trend forecasting, multi-factor correlation, and infrastructure health assessment are difficult to conduct effectively.

[0006] Therefore, integrating, managing, and sharing operations and maintenance knowledge to improve the efficiency and quality of operations and maintenance has become an essential and crucial step in enterprise group IT development. The creation of an operations and maintenance knowledge base is precisely aimed at addressing this challenge. However, during this process, the following challenges arose:

[0007] Knowledge entry and updating are cumbersome. Traditional operation and maintenance knowledge bases rely on manual operations, which are time-consuming, labor-intensive, and prone to errors. The rapid development of technology and the continuous emergence of new equipment make it difficult for the knowledge base to keep up with the pace of technological change, resulting in outdated or erroneous knowledge.

[0008] Knowledge retrieval efficiency is low. The retrieval function of traditional operation and maintenance knowledge base is usually based on keyword matching, which is easily affected by the personal experience and understanding of operation and maintenance personnel, resulting in inaccurate retrieval results or inability to find the required knowledge. At the same time, the documented storage method of knowledge lacks structured representation and association, making it difficult to perform in-depth retrieval and intelligent recommendations.

[0009] Knowledge sharing and inheritance are difficult. Traditional operations and maintenance knowledge bases are often private resources within the enterprise or even individuals, lacking mechanisms for exchanging and sharing external knowledge and experience. This limits operations and maintenance personnel's access to the latest technologies and solutions. Furthermore, the individual or team-based storage and management of knowledge hinders the effective inheritance and accumulation of knowledge.

[0010] Traditional O&M knowledge bases lack intelligent capabilities, such as automatic classification, intelligent recommendations, and automated question-and-answer (Q&A), making it more difficult for O&M personnel to solve complex problems. Furthermore, the lack of intelligent monitoring and early warning mechanisms makes it difficult to identify and warn of risks during O&M. Summary of the Invention

[0011] To overcome the problems existing in the related art, the disclosed embodiments of the present invention provide a method and system for automatically constructing a knowledge base in the field of IT operation and maintenance.

[0012] The technical solution is as follows: A method for automatically constructing an IT operation and maintenance domain knowledge base includes the following steps:

[0013] S1, builds the operation and maintenance knowledge model based on the RDF model;

[0014] S2, based on the constructed operation and maintenance knowledge model, builds a prompt template and uses a large language model combined with prompt words to parse and extract knowledge from the original document data;

[0015] S3, performs knowledge fusion on data with problems in knowledge extraction quality, knowledge duplication from different data sources, and association between knowledge;

[0016] S4, persistently stores the results of knowledge fusion and stores vector representations;

[0017] S5, for the data stored in vector representation, uses the graph path and vectorized hybrid retrieval method to integrate semantic relationships and weighted path relationships to obtain the information required for operation and maintenance knowledge retrieval.

[0018] In step S1, constructing an operation and maintenance knowledge model based on the RDF model includes:

[0019] S101, determining resources, where the resources are IT operation and maintenance objects, including servers, network security equipment, terminal devices, software applications, operation and maintenance tasks, etc.;

[0020] S102 defines attributes and relationships. Attributes include resource status, configuration, performance, and faults, and serve as attributes in the RDF model. Relationships include dependencies and connections between different resources. The RDF model is a standard that includes resources, attributes, and values. A triple (resource, attribute, value) is the smallest unit of RDF description. The operation and maintenance knowledge model is composed of countless triples.

[0021] S103, creating a data dictionary for defining metadata items for data modeling; the metadata items include a collection of classes and properties of object instances; properties can be further divided into attributes representing a class and relationships representing relationships between multiple classes; a unique URI is assigned to each resource, property, and relationship as a node in the RDF graph;

[0022] S104, constructing RDF elements, including: defining subjects (resources), selecting predicates (attributes, relationships), and determining objects (specific values, resources);

[0023] Based on the above steps S101 to S104, an operation and maintenance knowledge model is finally constructed to reflect the internal structure and logical relationship of the operation and maintenance domain knowledge, and to reflect the association between different operation and maintenance objects and the connection between various operation and maintenance operations.

[0024] Furthermore, after the operation and maintenance knowledge model is constructed, when new operation and maintenance knowledge or relationships appear, only new RDF classes or attributes are defined and integrated with the existing data model as attributes to complete the expansion of the operation and maintenance knowledge model.

[0025] The operation and maintenance knowledge model is mainly an IT operation and maintenance object data model, which is used to describe IT operation and maintenance objects and their related attributes, as well as fault types, and to show the association between IT operation and maintenance object instances and fault types.

[0026] In step S2, the large language model is used in combination with prompt words to parse and extract knowledge from the original document data, including:

[0027] S201, original document data preprocessing;

[0028] S202, knowledge extraction;

[0029] S203, knowledge relationship extraction.

[0030] In step S201, the original document data preprocessing includes:

[0031] S2011, converts documents in formats such as Word, PDF, and PPT into text format, removes irrelevant characters from the text, and standardizes the text format, setting uniform indentation and line breaks;

[0032] S2012, based on the logical structure and content of the original document, the text is segmented into different text units by identifying chapter marks, paragraph marks, indentation size, blank lines, and indent characters to determine segmentation boundaries;

[0033] S2013, identifies and extracts titles and subtitles based on text position, hierarchy, and indentation size; identifies lists based on text layout, alignment, and indentation patterns; and identifies table information based on row and column structure features;

[0034] In step S202, knowledge extraction includes:

[0035] S2021, design corresponding prompt word templates based on the operation and maintenance knowledge model and the specific content of the data.

[0036] S2022: Input the designed prompt words into the LLM according to the prompt word template, extract knowledge from the processed data according to the prompt words, and output it according to the data format requirements of the prompt word template; perform preliminary screening and sorting of the LLM output;

[0037] In step S203, knowledge relationship extraction includes limited domain relationship extraction and open domain relationship extraction, wherein limited domain relationship extraction includes: determining the existing relationship type based on the limited domain, and reflecting the relationship type in the operation and maintenance knowledge model; the relationship type includes the relationship between IT operation and maintenance objects, the relationship between IT operation and maintenance objects and fault types, and the relationship between faults and solutions.

[0038] In step S3, the knowledge fusion includes entity alignment and attribute alignment;

[0039] The entity alignment includes integrating data of entities that are the same or refer to the same entity into one entity object;

[0040] The attribute alignment includes using a word vector model to map words in attribute values to a vector space, measuring semantic similarity by calculating the cosine similarity between vectors, and determining attribute equivalence relationships in combination with dynamic threshold segmentation.

[0041] In step S4, persistent storage includes: storing the nodes and relationships in the operation and maintenance knowledge graph into a graph database for traversing the graph structure, locating relevant nodes and connections, obtaining complete knowledge information, and realizing rapid retrieval and traversal of knowledge;

[0042] The vector representation storage includes converting the nodes and relationships in the operation and maintenance knowledge graph into vector representations and storing them in the memory database.

[0043] In step S5, in the hybrid graph path and vectorized search method, the graph path search is used to find knowledge nodes and paths directly related to the query; the vectorized search is used to find knowledge nodes with semantic similarity to the query.

[0044] In step S5, the graph path and vectorized hybrid retrieval method includes assigning different weights to the graph path search results and the vectorized search results according to the task requirements to meet the task requirements and achieve the optimal fusion of semantic relations and weighted path relations.

[0045] Another object of the present invention is to provide a system for automatically constructing a knowledge base in the field of IT operation and maintenance, the system utilizing the method for automatically constructing a knowledge base in the field of IT operation and maintenance, the system comprising:

[0046] Operation and maintenance knowledge model construction module, which builds the operation and maintenance knowledge model based on the RDF model;

[0047] The knowledge extraction module builds prompt templates based on the established operation and maintenance knowledge model, and uses a large language model combined with prompt words to parse and extract knowledge from the original document data;

[0048] The knowledge fusion module is used to perform knowledge fusion on data with problems in knowledge extraction quality, knowledge duplication from different data sources, and knowledge association;

[0049] The knowledge storage module is used to persistently store the results of knowledge fusion and store vector representations;

[0050] The knowledge retrieval module is used to obtain the information required for operation and maintenance knowledge retrieval by using a hybrid retrieval method of graph path and vectorization to store data represented by vectors, integrating semantic relationships and weighted path relationships.

[0051] Combining all the above technical solutions, the beneficial effects of the present invention are as follows: The present invention proposes a method for constructing an intelligent operation and maintenance knowledge base based on artificial intelligence and big data technology. This method realizes the effective organization and structured representation of operation and maintenance knowledge by constructing an operation and maintenance knowledge data model based on RDF; utilizes a large language model for efficient knowledge extraction, which reduces the workload of manual entry while significantly reducing the error rate of knowledge processing and improving the efficiency and accuracy of knowledge updating; adopts a storage method that combines a graph database with an in-memory database to realize persistent storage and efficient retrieval of knowledge; and combines a large model to organize the knowledge in the knowledge base into valuable answers, providing intelligent support for operation and maintenance personnel, improving operation and maintenance efficiency and quality, and providing strong support for intelligent operation and maintenance.

[0052] The present invention provides a knowledge data model built on RDF, which enables various types of knowledge to be interconnected and integrated into a knowledge graph, greatly enhancing the relevance between knowledge and effectively making up for the defects of singleness and isolation in traditional knowledge representation methods. The present invention provides a vertical domain model - IT operation and maintenance field, which relies on the vertical domain model to efficiently extract knowledge, and refine a wide range of diverse raw data into high-quality knowledge, significantly improving the accuracy, integrity and comprehensiveness of knowledge, and laying a solid foundation for building a solid and reliable knowledge system. The present invention quickly converts document materials (operation and maintenance manuals, operation manuals, product white papers, technical parameters, etc.) into knowledge base data. The present invention improves the accuracy of knowledge retrieval and meets the needs of deep retrieval and intelligent recommendation. BRIEF DESCRIPTION OF THE DRAWINGS

[0053] The accompanying drawings, which are incorporated in and constitute a part of this specification, illustrate embodiments consistent with the present disclosure and, together with the description, serve to explain the principles of the present disclosure;

[0054] Figure 1 This is a flow chart of a method for automatically constructing an IT operation and maintenance domain knowledge base provided by an embodiment of the present invention;

[0055] Figure 2 This is a schematic diagram of a method for automatically constructing an IT operation and maintenance domain knowledge base provided by an embodiment of the present invention;

[0056] Figure 3 This is a schematic diagram of a system for automatically constructing an IT operation and maintenance knowledge base provided by an embodiment of the present invention;

[0057] In the figure: 1. Operation and maintenance knowledge model construction module; 2. Knowledge extraction module; 3. Knowledge fusion module; 4. Knowledge storage module; 5. Knowledge retrieval module. DETAILED DESCRIPTION

[0058] To make the above-mentioned objects, features, and advantages of the present invention more readily apparent, specific embodiments of the present invention are described in detail below with reference to the accompanying drawings. The following description sets forth numerous specific details to facilitate a full understanding of the present invention. However, the present invention can be implemented in many other ways than those described herein, and those skilled in the art may make similar modifications without departing from the scope of the present invention. Therefore, the present invention is not limited to the specific embodiments disclosed below.

[0059] The innovation of the present invention lies in: proposing a method for constructing an intelligent operation and maintenance knowledge base based on artificial intelligence and big data technology, realizing effective organization and structured representation of operation and maintenance knowledge by constructing an operation and maintenance knowledge data model based on RDF; utilizing a large language model for efficient knowledge extraction, reducing the workload of manual entry while significantly reducing the error rate of knowledge processing, and improving the efficiency and accuracy of knowledge updating; adopting a storage method combining a graph database and an in-memory database, realizing persistent storage and efficient retrieval of knowledge; and combining a large model to organize the knowledge in the knowledge base into valuable answers, providing intelligent support for operation and maintenance personnel, improving operation and maintenance efficiency and quality, and providing strong support for intelligent operation and maintenance.

[0060] In Example 1, the present invention aims to utilize advanced artificial intelligence technologies (knowledge graphs, natural language processing technologies) to construct a new vertical domain model—the IT operations and maintenance macro model—through in-depth analysis of data extraction, organization, and storage methods in the IT operations and maintenance field. Based on this model, an intelligent operations and maintenance knowledge base is constructed, providing an innovative solution for the creation of an operations and maintenance knowledge base, as well as innovative applications of artificial intelligence technologies in the vertical field of IT operations and maintenance. This mainly includes:

[0061] To organize knowledge, the present invention defines a knowledge data model and uses the model to define knowledge; utilizes natural language preprocessing technology to parse and extract the original knowledge data in segments according to the knowledge data model to form knowledge; combines knowledge graph technology to establish the relationship between knowledge according to the characteristics of IT operation and maintenance.

[0062] Regarding knowledge storage, this paper proposes a data storage method that supports hybrid graph path and vectorized retrieval. This method ensures both persistent knowledge storage and efficient and accurate knowledge retrieval. It can retrieve not only directly semantically related knowledge but also indirectly semantically related knowledge. This improves the accuracy of retrieval results while meeting the needs of deep retrieval and intelligent recommendations.

[0063] Nouns related to the present invention include:

[0064] "RDF": The RDF data model is similar to classic conceptual modeling approaches (such as entity-relationship or class diagrams) and is based on the idea of expressing statements about resources in the form of subject-verb-object triples. The subject represents the resource, the predicate represents a feature or aspect of the resource, and the object represents the relationship between the subject and the object. RDF is an abstract model with several serialization formats (essentially specialized file formats), and the specific encoding of resources or triples may vary between formats. A collection of RDF statements essentially represents a labeled directed graph, which makes the RDF data model more suitable for knowledge representation than other relational or ontology models.

[0065] "Knowledge Graph": A knowledge graph is essentially a semantic network and a structured semantic knowledge base. It consists of triplets such as "entity-relationship-entity" or "entity-attribute-attribute value" and is presented as a graph. Nodes in the graph represent concepts or entities, and edges represent relationships or attributes between things. The knowledge source for constructing a knowledge graph comes from relevant structured, semi-structured, and unstructured data. Knowledge graphs are divided into general knowledge graphs and domain knowledge graphs. General knowledge graphs involve common sense knowledge and demonstrate the breadth of knowledge; domain knowledge graphs involve specific domain knowledge and reflect the depth of knowledge. The operations knowledge graph is a typical domain knowledge graph. The logical structure of a knowledge graph consists of two layers: the schema layer and the data layer. There are two approaches to constructing it: top-down (schema layer first, then data layer) and bottom-up (data layer first, then schema layer). Domain knowledge graphs are mostly constructed using the top-down approach, primarily because the concepts, definitions, and frameworks involved in domain knowledge graphs have become stable empirical models through long-term accumulation and repeated practice.

[0066] like Figure 1 As shown, the method for automatically constructing an IT operation and maintenance domain knowledge base provided by an embodiment of the present invention includes:

[0067] S1, builds the operation and maintenance knowledge model based on the RDF model;

[0068] In the intelligent operation and maintenance knowledge base, the IT operation and maintenance knowledge graph is a typical domain knowledge graph. The effective organization and structured representation of operation and maintenance knowledge are the basis for achieving efficient question and answer. Taking the objects of IT operation and maintenance as the entry point for knowledge graph construction, the present invention uses RDF (Resource Description Framework) as the core framework of the operation and maintenance knowledge data model to create an operation and maintenance knowledge model. RDF provides a flexible and extensible way to describe network resources and the relationships between them, and is very suitable for representing complex knowledge systems in the operation and maintenance field. The goals and scope of the IT operation and maintenance knowledge model cover the categories of IT operation and maintenance objects, the required functional characteristics (such as performance, reliability, security, etc.), and how the model will be used (such as for decision support, troubleshooting, performance optimization, etc.).

[0069] Specifically include:

[0070] S101, Determine Resources. First, identify the resources that need to be described. In the context of the IT operations knowledge model, these resources are IT operations objects, including servers, network security devices, software applications, operations tasks, etc.

[0071] S102: Define attributes and relationships. Based on the characteristics of the resources, define a vocabulary to describe attributes such as resource status, configuration, performance, and faults. These attributes will serve as edges or relationships in the RDF graph. At the same time, determine any relationships between different resources, such as dependencies and connections.

[0072] S103, create a data dictionary. Before building the RDF graph, create a data dictionary to define the metadata items for data modeling. This includes a collection of object instances (classes) and properties. Properties can be further divided into attributes representing classes and relationships between multiple classes. Assign a unique URI (Uniform Resource Identifier) to each resource, property, and relationship. These URIs will serve as nodes in the RDF graph.

[0073] S104: Construct RDF elements and use the RDF graph data model to represent these triples.

[0074] (1) Define the subject: In the RDF model, the subject is the central entity involved in the declaration or description, which can be concrete or abstract. For the IT operation and maintenance knowledge model, the subject includes the IT operation and maintenance object itself, the hardware components of the operation and maintenance object (such as CPU, memory, hard disk), software components (such as operating system, database, middleware), configuration items of the operation and maintenance object, and operation and maintenance tasks (performance monitoring, troubleshooting, etc.).

[0075] (2) Select predicates: Predicates describe the relationship between the subject and the object. In the IT operation and maintenance knowledge model, predicates are vocabulary such as device attributes (such as model, manufacturer), configuration parameters (such as IP address, port number), performance indicators (such as CPU usage, memory usage), and fault types (such as hardware fault, software fault). When defining the model, the present invention embodies predicates in the form of attributes.

[0076] (3) Determine the object: The object is the entity or value pointed to by the predicate. In the IT operation and maintenance knowledge model, the object is a specific value (such as a number or string), other resources (such as the URI of another device or software system), or a blank node (indicating that there is no specific object, but the relationship exists). In the model, the object is specifically the value of an attribute.

[0077] Illustratively, the present invention defines a server data model, which describes servers and their related attributes, as well as failure types, and shows how to associate server instances with failure types.

[0078] The server data model includes:

[0079] 1) Server class (ex: Server): represents the abstract concept of a server.

[0080] 2) Attributes include:

[0081] ex: hasModelName: Server model, such as "Dell PowerEdge R740".

[0082] ex: hasManufacturer: Manufacturer, such as "Dell".

[0083] ex: hasIPAddress: The IP address of the server, for example "192.168.1.10".

[0084] ex: hasCPUCount: number of CPUs, for example 2.

[0085] ex: hasMemorySize: memory capacity, for example "32GB".

[0086] ex: isRunning: server status, indicating whether the server is running, for example, true.

[0087] 3) Fault type class (ex: FaultType): represents the abstract concept of fault type.

[0088] 4) Subcategories include: ex: HardwareFault: hardware failure; ex: SoftwareFault: software failure.

[0089] 5) Relationships include:

[0090] ex: hasFault: represents the relationship between a server instance and a fault type.

[0091] The resulting knowledge model, built through the above process, is a data model specifically tailored for operations knowledge. This model reflects the inherent structure and logical relationships within the domain. For example, it can capture the connections between different operations devices (such as servers and switches) and their connections to various operations (such as configuration and troubleshooting).

[0092] S105, Expansion of the Operation and Maintenance Knowledge Model. The flexibility and scalability of the RDF framework enable the Operation and Maintenance Knowledge Model to continuously expand as the field of Operation and Maintenance evolves. When new Operation and Maintenance knowledge or relationships emerge, the present invention simply defines new RDF classes or attributes and integrates them with the existing data model as attributes, completing the expansion. This scalability ensures that the Operation and Maintenance Knowledge Model can continuously adapt to changes in the field of Operation and Maintenance.

[0093] S2, based on the constructed operation and maintenance knowledge model, builds a prompt template and uses a large language model combined with prompt words to parse and extract knowledge from the original document data;

[0094] Knowledge discovery is a key step in achieving knowledge organization. The present invention adopts a multi-step knowledge discovery process, including data preprocessing, feature extraction, operation and maintenance knowledge model training and knowledge extraction.

[0095] Specifically include:

[0096] S201, preprocessing of original document data.

[0097] In S2011, use tools to convert documents in word, pdf, ppt and other formats into text formats that can be processed by the large model, remove irrelevant characters in the text such as headers, footers, advertisements, watermarks, etc., and standardize the text format with unified indentation and line breaks.

[0098] S2012, based on the logical structure (such as chapters, paragraphs) and content of the original document, the text is segmented by identifying chapter marks, paragraph marks, indentation size, blank lines, and indentation characters to determine the segmentation boundaries, and finally segmented into different text units.

[0099] S2013, identifies and extracts titles and subtitles based on text position, hierarchy, and indentation size; identifies lists based on text layout, alignment, and indentation patterns; and identifies table information based on row and column structure features.

[0100] S202, knowledge extraction.

[0101] S2021: Design corresponding prompt word templates based on the operation and maintenance knowledge model and the specific content of the data. The prompt words should clearly and accurately describe the type of knowledge to be extracted.

[0102] This invention uses product manuals as an example to extract basic device information. Within the O&M knowledge model, this invention requires extracting information such as the server's manufacturer, model, CPU model, number of CPUs, memory capacity, and dimensions from the product manual. Based on the O&M knowledge model, a prompt word template is constructed to extract the required information from the server's product manual and output it according to the O&M knowledge model.

[0103] The teleprompter template extracts server-related information from the operation manual according to the following RDF model structure and outputs it in the form of RDF triples:

[0104] (1) Server class (ex: Server) and its properties:

[0105] Model (ex: hasModelName);

[0106] Manufacturer (ex: hasManufacturer);

[0107] CPU model (ex: hasCPUType);

[0108] Number of CPUs (ex: hasCPUCount);

[0109] Memory capacity (e.g. hasMemorySize);

[0110] size (ex: hasSize);

[0111] (2) Fault type class (ex: FaultType) and its subclasses: hardware fault (ex: HardwareFault), software fault (ex: SoftwareFault);

[0112] (3) Relationship between server and fault type (ex: hasFault): Please indicate what type of fault occurred in the server instance.

[0113] S2022: Input the designed prompt words into the LLM using the prompt word template described above. The large model extracts knowledge from the processed data based on the prompt words and outputs it in the data format required by the prompt word template. Preliminary screening and organization of the LLM output ensures that the extracted knowledge meets the requirements. This primarily involves named entity recognition, such as device name, model name, software version, fault type, and solution.

[0114] S203, knowledge relationship extraction.

[0115] There are two ways to extract relationships: limited domain relationship extraction and open domain relationship extraction. The limited domain is the IT operation and maintenance domain. The relationship types that should exist are determined based on the limited domain, and the relationship types are reflected in the operation and maintenance knowledge model. When designing the operation and maintenance knowledge model, the present invention establishes relationships based on the principle that the subject of a certain operation and maintenance knowledge model is the object of another operation and maintenance knowledge model. Therefore, the relationship between entities has been designed and reflected in the operation and maintenance knowledge model, such as the relationship between servers and hardware failures, the relationship between servers and failure types, and the relationship between failures and solutions. The process of constructing the operation and maintenance knowledge model is also a reflection of the limited domain relationship extraction process.

[0116] In the process of identifying knowledge and mapping it to the operation and maintenance knowledge model, the relationship between knowledge has also been established. Based on the professional knowledge in the operation and maintenance field, the relationship between subjects is mainly as follows:

[0117] Hardware-software relationship: Establish the relationship between them based on the software installed on the hardware.

[0118] Hardware-to-hardware relationships: Establish relationships between devices based on how hardware relates to other hardware.

[0119] Hardware-fault relationship: Establishes a relationship between a device and its possible faults based on the device type and fault type.

[0120] Fault-Solution Relationship: Establish a relationship between a fault and its corresponding solution based on the fault description and solution content.

[0121] Other relationships: Based on actual needs, you can also establish dependency relationships between devices, causal relationships between faults, etc.

[0122] By establishing these relationships, the present invention can construct an operation and maintenance knowledge graph containing multi-dimensional information such as hardware, software, faults, solutions, etc., providing support for subsequent knowledge retrieval and question-answering.

[0123] S3, performs knowledge fusion on data with problems in knowledge extraction quality, knowledge duplication from different data sources, and association between knowledge;

[0124] Entity alignment: Integrate data for entities that are the same or refer to the same entity. For example, if a host and a server are represented as the same entity object, the data will be integrated into one entity object server. Manual confirmation is required here.

[0125] Attribute alignment: Different data sources may record entity attributes differently. We use word vector models (such as Word2Vec and BERT) to map the words in the attribute values into a vector space. We then calculate the cosine similarity between the vectors to measure their semantic similarity. If any attribute or attribute value is found to be partially redundant, its weight is reduced when calculating the similarity. The attribute or attribute value similarity calculation is then completed, and dynamic threshold segmentation is used to determine equivalence relationships.

[0126] After completion, implicit knowledge can be further mined based on the existing knowledge base, thereby enriching and expanding the knowledge base.

[0127] S4, persistently stores the results of knowledge fusion and stores vector representations;

[0128] S401, persistent storage. To achieve persistent storage of knowledge, the present invention employs a graph database for storage. Graph databases excel at handling complex relationship structures between entities and are well-suited for storing knowledge systems such as the operation and maintenance knowledge graph, which contain numerous entities and relationships. The present invention stores the nodes and relationships in the operation and maintenance knowledge graph in a graph database. The graph database can efficiently traverse the graph structure, quickly locate relevant nodes and connections, obtain complete knowledge information, and achieve rapid retrieval and traversal of knowledge.

[0129] S402, vector representation storage. To further improve the efficiency of knowledge retrieval, the present invention also uses an in-memory database for storage, which supports efficient text search and vectorized retrieval. The present invention converts the nodes and relationships in the operation and maintenance knowledge graph into vector representations and stores them in the in-memory database, namely the vector database.

[0130] S5, for the data stored in vector representation, uses the graph path and vectorized hybrid retrieval method to integrate semantic relationships and weighted path relationships to obtain the information required for operation and maintenance knowledge retrieval.

[0131] Hybrid graph and vector search. This invention combines graph path search and vectorized search methods during the search process. Traditional vector search, such as searching for switch maintenance knowledge, only retrieves content directly related to the semantics of "maintenance" and "switch." However, in actual operations and maintenance scenarios, it is often necessary to consider data indirectly related to semantic logic.

[0132] The in-memory database in the present invention solves this problem through the ability of graph path and vectorized fusion search. Through graph path retrieval, the present invention can find knowledge nodes and paths directly related to the query terms; through vectorized retrieval, the present invention can find knowledge nodes with semantic similarity to the query terms.

[0133] Graph path search involves traversing adjacent nodes in the graph database's graph structure based on the entities and relationships involved in the query, starting with the query starting entity and following predefined relationship types. For each node, check whether its relationship with the next node meets the query requirements. By iterating this process, possible paths are gradually explored while avoiding repeated visits to nodes that have already been searched. Paths are continuously expanded along relationships that meet the requirements until the target entity is found or all possible paths have been traversed. This results in a path from the starting entity to the target entity that satisfies the entities and relationships involved in the query, thus completing the path search in the graph structure.

[0134] The vectorized search includes: searching in the vector space to meet the semantic similarity requirements in the query; if the query requires finding other entities that are semantically similar to a certain entity, first obtain the vector representation of the query entity and the vector representation of each of the searchable nodes in the graph. For each searchable node, according to the cosine similarity calculation formula, the dot product of the query entity vector and the node vector is calculated divided by the product of the module lengths of the two vectors. Specifically, the sum of the products of the corresponding dimension values of the vectors is calculated as the dot product result, and then the module lengths of the query entity vector and the node vector are respectively calculated (obtained by taking the square root of the sum of the squares of the values of each dimension), and finally the dot product is divided by the product of the two module lengths to obtain the cosine similarity between the two, so as to measure the distance feature between the query entity and each searchable node, thereby reflecting the degree of similarity between them, and calculating the distance measure (such as cosine similarity) between the vectorized query entity and all searchable nodes (vectorized) in the graph; according to the set similarity threshold, the nodes that meet the requirements are screened out.

[0135] The hybrid retrieval strategy of graph path retrieval and vectorization, i.e., the fusion strategy, includes:

[0136] Weighted fusion assigns different weights to graph path search results and vectorized search results based on task requirements. For example, if the accuracy of relationships within the graph structure is emphasized, the graph path search results will be given a higher weight; if semantic relevance is emphasized, the vectorized search results will be given a higher weight. For example, for an information retrieval task, analysis found that a 60% structural relationship is more conducive to result accuracy, while a 40% semantic similarity is sufficient. Therefore, a weight of 0.6 is assigned to the graph path search results, and a weight of 0.4 is assigned to the vectorized search results. In actual applications, weights may also be continuously adjusted based on new feedback and data to ensure that the weight allocation can best meet the task requirements, thereby achieving the optimal fusion effect.

[0137] The hybrid retrieval method of graph path retrieval and vectorized retrieval in the present application takes into account both semantic relationships and relationships between weighted paths, thereby improving the accuracy and comprehensiveness of knowledge retrieval. For example, when querying switch troubleshooting knowledge, in addition to obtaining knowledge directly related to switch fault diagnosis, it is also possible to retrieve relevant knowledge about the possible impact of other network devices connected to it on switch failures. In other words, it not only takes into account situations that are directly semantically related, but also mines knowledge that is indirectly related to semantic logic along the graph path. This hybrid retrieval of graph path and vectorization combines semantic relationships and weighted path relationships, and can provide the required information more comprehensively and accurately in operation and maintenance knowledge retrieval.

[0138] Example 2, as Figure 2As shown, the method principle of automatically constructing the IT operation and maintenance field knowledge base provided by the embodiment of the present invention includes: knowledge modeling, knowledge acquisition and extraction, knowledge fusion, knowledge storage, and hybrid retrieval of graph path retrieval and vectorization;

[0139] Specifically include:

[0140] Step 1: Create an IT operation and maintenance data model based on the RDF model;

[0141] Step 2: Construct a prompt word template based on the IT operation and maintenance data model. After data preprocessing of the original document (PDF / Word / PPT), perform knowledge extraction on the received prompt word template, including: LLM-based knowledge extraction, output according to the RDF data model, and manual verification.

[0142] Step 3: After knowledge extraction, knowledge fusion is performed, including entity alignment, attribute alignment, and knowledge reasoning;

[0143] Step 4: Store the knowledge representation after knowledge fusion in the graph database, and store the data after knowledge vectorization of the knowledge representation in the vector database.

[0144] Later, the stored data is subjected to hybrid retrieval of graph path retrieval and vectorization.

[0145] Example 3, as another implementation of the present invention, the present invention provides an application of a method for automatically constructing an IT operation and maintenance field knowledge base.

[0146] When using the automatic construction of the IT operation and maintenance field knowledge base, it is combined with LLM to organize answers. When the operation and maintenance personnel ask a question, such as "How to optimize the performance of the server in the network while ensuring data security", the large model first parses the semantics of the question, and then searches for relevant knowledge in the knowledge base using a hybrid retrieval mechanism of graphs and vectors. Relevant knowledge nodes such as server performance optimization (such as knowledge on hardware upgrades, software tuning, etc.) and data security (such as knowledge on firewall settings, user rights management, etc.) in the knowledge graph provided by the knowledge base will be retrieved. Based on this retrieved knowledge, the large model organizes answers in a certain logical order. For example, it first introduces the performance improvement brought by hardware upgrades and the possible data security risks, and then explains how to balance performance optimization and data security requirements through firewall settings. This can provide operation and maintenance personnel with clear, comprehensive and targeted answers.

[0147] Example 4, as Figure 3 As shown, an embodiment of the present invention provides a system for automatically building an IT operation and maintenance domain knowledge base, including:

[0148] Operation and maintenance knowledge model construction module 1, constructs the operation and maintenance knowledge model based on the RDF model;

[0149] Knowledge extraction module 2, based on the constructed operation and maintenance knowledge model, builds a prompt template and uses a large language model combined with prompt words to parse and extract knowledge from the original document data;

[0150] Knowledge fusion module 3 is used to perform knowledge fusion on data with problems in knowledge extraction quality, knowledge duplication from different data sources, and knowledge association;

[0151] The knowledge storage module 4 is used for persistent storage and vector representation storage of the results of knowledge fusion;

[0152] The knowledge retrieval module 5 is used to obtain the information required for operation and maintenance knowledge retrieval by using a hybrid retrieval method of graph path and vectorization to store data represented by vectors, integrating semantic relationships and weighted path relationships.

[0153] The present invention improves the efficiency and quality of operation and maintenance: by constructing an intelligent operation and maintenance knowledge base, operation and maintenance personnel can find the required knowledge and answers more quickly, thereby improving the efficiency and quality of operation and maintenance. Reduce the workload of manual input and lower the error rate of knowledge processing: by using a large language model for knowledge extraction, the workload of manual input is reduced while significantly reducing the error rate of knowledge processing, thereby improving the efficiency and accuracy of knowledge updating. Improve the accuracy of knowledge retrieval: by adopting a storage method that combines a graph database with an in-memory database, persistent storage and efficient retrieval of knowledge are achieved, thereby improving the accuracy of knowledge retrieval. Provide intelligent support: by combining a large model, the knowledge in the knowledge base is organized into valuable answers, providing intelligent support for operation and maintenance personnel. Improve the relevance and integration of knowledge: the present invention provides a vertical domain model - the IT operation and maintenance field, which relies on this vertical domain model to efficiently extract knowledge and refine a wide range of diverse raw data into high-quality knowledge. Quickly convert document materials into knowledge base data: the present invention quickly converts document materials (operation and maintenance manuals, operation manuals, white papers, etc.) into knowledge base data, thereby improving the efficiency of knowledge base construction. Meeting the needs of deep retrieval and intelligent recommendation: The present invention improves the accuracy of knowledge retrieval, meets the needs of deep retrieval and intelligent recommendation, and improves user experience.

[0154] The present invention provides a knowledge data model built based on RDF, which solves the problems existing in the traditional operation and maintenance knowledge base. The present invention provides a vertical field model - IT operation and maintenance field, and relies on this vertical field model to efficiently extract knowledge; it can efficiently and accurately operate and maintain knowledge. It improves the accuracy of knowledge retrieval and meets the needs of deep retrieval and intelligent recommendation. The method of the present invention for quickly converting document materials into knowledge base data improves the efficiency of knowledge base construction and expands the volume of the operation and maintenance knowledge base. The knowledge data model constructed based on RDF in the present invention improves the relevance and integration between knowledge. The present invention improves the accuracy of knowledge retrieval, meets the needs of deep retrieval and intelligent recommendation, and improves user experience.

[0155] The method of the present invention for quickly converting document data into knowledge base data improves the efficiency of knowledge base construction. The present invention uses a large language model to extract knowledge, which reduces the workload of manual input while significantly reducing the error rate of knowledge processing, and improves the efficiency and accuracy of knowledge updating. The knowledge data model constructed by the present invention based on RDF improves the relevance and integration between knowledge. The present invention improves the accuracy of knowledge retrieval, meets the needs of deep retrieval and intelligent recommendation, and improves user experience. By constructing an operation and maintenance knowledge data model based on RDF, the present invention realizes the interconnection and integration of various types of knowledge and forms a knowledge graph in the operation and maintenance field. This solves the defects of singleness and isolation in traditional knowledge representation methods and enhances the relevance between knowledge. The present invention proposes an efficient knowledge extraction method based on a large language model, which can automatically parse and extract knowledge from original documents, reduce the workload of manual input, and improve the efficiency and accuracy of knowledge updating. The present invention adopts a storage method that combines a graph database with an in-memory database and implements a hybrid retrieval method of graph path and vectorization, which not only ensures the persistent storage of knowledge but also improves the efficiency and accuracy of knowledge retrieval. This method can retrieve knowledge with direct and indirect semantic associations, meeting the needs of deep retrieval and intelligent recommendation.

[0156] The intelligent operation and maintenance knowledge base mentioned in the present invention combines natural language preprocessing technology to organize and answer knowledge, realizes functions such as automatic classification, intelligent recommendation, and automatic question and answer, provides intelligent support for operation and maintenance personnel, and reduces the difficulty of solving complex problems.

[0157] This method can accurately identify knowledge triples in operation and maintenance documents and construct a complete operation and maintenance knowledge graph. At the same time, this method also has efficient storage and retrieval capabilities, enabling fast and accurate answers to operation and maintenance questions.

[0158] This invention, through in-depth research on the extraction, organization and storage methods of raw data related to IT operation and maintenance, constructs a large model of a vertical field - the IT operation and maintenance field, and uses this model to create an intelligent operation and maintenance knowledge base. The knowledge base constructs a knowledge data model based on RDF, uses a large language model for efficient knowledge extraction, and adopts a storage method that combines a graph database with an in-memory database to achieve persistent storage and efficient retrieval of knowledge. During the use phase, the knowledge in the knowledge base is organized into valuable answers in combination with the large model. Through this method, the article successfully solves the problems existing in traditional operation and maintenance knowledge bases in knowledge representation, entry and update, retrieval efficiency, sharing and inheritance, and intelligent support, providing strong support for intelligent operation and maintenance. These innovative methods help to improve operation and maintenance efficiency and quality, and further promote the development of the field of intelligent operation and maintenance.

[0159] The above description is only a preferred specific implementation method of the present invention, but the scope of protection of the present invention is not limited thereto. Any modifications, equivalent substitutions and improvements made by any technician familiar with this technical field within the technical scope disclosed by the present invention and within the spirit and principles of the present invention should be covered by the scope of protection of the present invention.

Claims

1. A method for automatically constructing an IT operation and maintenance domain knowledge base, characterized in that: The method comprises the following steps: S1, builds the operation and maintenance knowledge model based on the RDF model; S2, based on the constructed operation and maintenance knowledge model, builds a prompt template and uses a large language model combined with prompt words to parse and extract knowledge from the original document data; S3, performs knowledge fusion on data with problems in knowledge extraction quality, knowledge duplication from different data sources, and association between knowledge; S4, persistently stores the results of knowledge fusion and stores vector representations; S5, for the data stored in vector representation, uses the graph path and vectorized hybrid retrieval method to integrate semantic relationships and weighted path relationships to obtain the information required for operation and maintenance knowledge retrieval.

2. The method for automatically constructing an IT operation and maintenance domain knowledge base according to claim 1, characterized in that: In step S1, an operation and maintenance knowledge model is constructed based on the RDF model, including: S101, determining resources, where the resources are IT operation and maintenance objects, including servers, network security devices, terminal devices, software applications, and operation and maintenance tasks; S102, defining attributes and relationships, wherein the attributes include: resource status, configuration, performance, and fault, and are used as attributes in the RDF model; the relationships include: dependency relationships and connection relationships between different resources; S103, creating a data dictionary for defining metadata items for data modeling; the metadata items include a collection of classes and properties of object instances; properties are divided into attributes representing classes and relationships representing relationships between multiple classes; a unique URI is assigned to each resource, property, and relationship as a node in the RDF graph; S104, constructing RDF elements, including: defining the subject, selecting the predicate, and determining the object; Finally, an operation and maintenance knowledge model is constructed, and an object data model is constructed based on IT operation and maintenance objects. This model is used to describe IT operation and maintenance objects and their related attributes, as well as the fault types of operation and maintenance objects, and to show the association between IT operation and maintenance object instances and fault types. It is used to reflect the inherent structure and logical relationship of operation and maintenance domain knowledge, and to reflect the association between different operation and maintenance objects and the connection between them and various operation and maintenance operations.

3. The method for automatically constructing an IT operation and maintenance domain knowledge base according to claim 2, characterized in that: After the operation and maintenance knowledge model is constructed, when new operation and maintenance knowledge or relationships appear, new RDF classes or attributes are defined and used as attributes or relationships of the existing data model, integrated with the existing data model to complete the expansion of the operation and maintenance knowledge model.

4. The method for automatically constructing an IT operation and maintenance domain knowledge base according to claim 1, characterized in that: In step S2, the large language model is used in combination with prompt words to parse and extract knowledge from the original document data, including: S201, original document data preprocessing; S202, knowledge extraction; S203, knowledge relationship extraction, including limited domain relationship extraction and open domain relationship extraction; limited domain relationship extraction includes: determining the existing relationship types based on the limited domain, and reflecting the relationship types in the operation and maintenance knowledge model; relationship types include the relationship between IT operation and maintenance objects, the relationship between IT operation and maintenance objects and fault types, and the relationship between faults and solutions.

5. The method for automatically constructing an IT operation and maintenance domain knowledge base according to claim 4, characterized in that: In step S201, the original document data is pre-processed, including: S2011, convert documents in word, pdf, and ppt formats into text format, remove irrelevant characters in the text, and standardize the text format, setting unified indentation and line breaks; S2012, based on the logical structure and content of the original document, the text is segmented into different text units by identifying chapter marks, paragraph marks, indentation size, blank lines, and indent characters to determine segmentation boundaries; S2013, identifies and extracts titles and subtitles based on text position, hierarchy, and indentation size; identifies lists based on text layout, alignment, and indentation rules; and identifies table information based on row and column structure features.

6. The method for automatically constructing an IT operation and maintenance domain knowledge base according to claim 4, characterized in that: In step S202, knowledge extraction includes: S2021: Design corresponding prompt word templates based on the operation and maintenance knowledge model and the specific content of the data; S2022: Input the designed prompt words into the LLM according to the prompt word template, extract knowledge from the processed data according to the prompt words, and output it according to the data format requirements of the prompt word template, and preliminarily screen and organize the LLM output.

7. The method for automatically constructing an IT operation and maintenance domain knowledge base according to claim 1, characterized in that: In step S3, knowledge fusion includes: entity alignment and attribute alignment; Entity alignment includes: integrating data of entities that are the same or refer to the same entity into one entity object; Attribute alignment includes: using the word vector model to map the words in the attribute value to the vector space, measuring the semantic similarity by calculating the cosine similarity between the vectors, and combining dynamic threshold segmentation to determine the attribute equivalence relationship.

8. The method for automatically constructing an IT operation and maintenance domain knowledge base according to claim 1, characterized in that: In step S4, persistent storage includes: storing the nodes and relationships in the operation and maintenance knowledge graph in a graph database for traversing the graph structure, locating relevant nodes and connections, obtaining complete knowledge information, and realizing rapid retrieval and traversal of knowledge; The vector representation storage includes converting the nodes and relationships in the operation and maintenance knowledge graph into vector representations and storing them in the memory database.

9. The method for automatically constructing an IT operation and maintenance domain knowledge base according to claim 1, characterized in that: In step S5, in the graph path and vectorized hybrid retrieval method, the knowledge nodes and paths directly related to the query term are found through graph path retrieval; Find knowledge nodes with similar semantics to the query term through vectorized retrieval; According to the task requirements, different weights are assigned to the graph path search results and vectorized search results to meet the task requirements and achieve the optimal fusion of semantic relationships and weighted path relationships.

10. A system for automatically building a knowledge base in the field of IT operation and maintenance, characterized by: The system is implemented by the method for automatically constructing an IT operation and maintenance domain knowledge base according to any one of claims 1 to 9, and the system includes: Operation and maintenance knowledge model construction module (1), constructs the operation and maintenance knowledge model based on the RDF model; The knowledge extraction module (2) builds a prompt template based on the constructed operation and maintenance knowledge model, and uses a large language model combined with prompt words to parse and extract knowledge from the original document data; The knowledge fusion module (3) is used to perform knowledge fusion on data with problems in knowledge extraction quality, knowledge duplication from different data sources, and association between knowledge; The knowledge storage module (4) is used for persistent storage and vector representation storage of the results of knowledge fusion; The knowledge retrieval module (5) is used to obtain the information required for operation and maintenance knowledge retrieval by using a graph path and vectorized hybrid retrieval method to integrate semantic relationships and weighted path relationships for the data stored in vector representation.

Images