CN-122019733-A - Intelligent device operation and maintenance question-answering method and system based on knowledge graph context fusion

CN122019733ACN 122019733 ACN122019733 ACN 122019733ACN-122019733-A

Abstract

The invention discloses a device operation and maintenance intelligent question-answering method and system based on knowledge graph context fusion, which comprises the steps of firstly constructing a knowledge text data set in the field of industrial device operation and maintenance, and acquiring an operation and maintenance knowledge graph of the industrial equipment, and carrying out community division on the knowledge graph to acquire a knowledge community. And secondly, respectively carrying out vectorization representation on the problems input by the user and the knowledge communities. And then calculating the diversity association relation between the problem vector and the community vector, generating similarity credibility distribution reflecting different association degrees, and carrying out adaptability adjustment through a community credibility discount strategy. And finally, fusing the similarity reliability distribution subjected to the adaptive adjustment, constructing a knowledge graph context, inputting the knowledge graph context into an industrial equipment operation and maintenance large language model, and generating answers of industrial equipment operation and maintenance questions. The invention improves the reliability of the constructed knowledge graph context and provides an efficient and decision-making credible intelligent device operation and maintenance question-answering method for the operation and maintenance of the industrial device.

Inventors

XU XIAOBIN
GE CHENGXUAN
HE HONG
ZHANG ZHENJIE
ZHANG ZEHUI
Gao Zhangnan
WANG XIAOFENG

Assignees

杭州电子科技大学

Dates

Publication Date: 20260512
Application Date: 20260413

Claims (9)

1. The intelligent equipment operation and maintenance question-answering method based on knowledge graph context fusion is characterized by comprising the following steps of: S1, constructing a knowledge text data set in the field of operation and maintenance of industrial equipment; s2, constructing an industrial equipment operation and maintenance knowledge graph based on a knowledge text data set, and carrying out community division on the knowledge graph to obtain a plurality of knowledge communities representing knowledge units with different operation and maintenance; s3, vectorizing the questions and knowledge communities input by the user to obtain question vectors and community vectors; S4, calculating the diversity association relation between the problem vector and the community vector through the differentiated similarity measurement, and generating similarity reliability distribution reflecting different association degrees; s5, adaptively adjusting similarity reliability distribution through a community reliability discount strategy; And S6, fusing the similarity reliability distribution subjected to the adaptive adjustment, constructing a knowledge graph context, inputting the knowledge graph context into an industrial equipment operation and maintenance large language model, and generating answers of industrial equipment operation and maintenance questions.
2. The intelligent query and answer method for equipment operation and maintenance based on knowledge-graph context fusion according to claim 1, wherein the specific implementation process of the step S1 is as follows: S1.1, uniformly collecting unstructured or semi-structured operation and maintenance data of different sources through a document analysis and text extraction technology, wherein original text data is obtained from an overhaul record database, a fault report document library and an equipment operation and maintenance standard file which are related to the operation and maintenance of industrial equipment; S1.2, cleaning, segmenting and semantically standardizing original text data, eliminating noise content and unifying professional term mapping to obtain a standardized operation and maintenance knowledge text data set; S1.3, knowledge items are extracted from the operation and maintenance knowledge text data set, and an industrial equipment operation and maintenance question and answer sample set containing a plurality of question and answer pairs is constructed.
3. The intelligent query and answer method for equipment operation and maintenance based on knowledge-graph context fusion according to claim 1, wherein the specific implementation process of the step S2 is as follows: S2.1, carrying out text unit division on the knowledge text data set of the operation and maintenance of the industrial equipment, which is obtained in the S1, identifying equipment objects, fault phenomena, fault parts and disposal measures from the knowledge text data set as entity nodes based on a named entity identification and relation extraction technology, and simultaneously identifying causal association relations and disposal dependency relations among the entity nodes to construct structured graph data so as to obtain an operation and maintenance knowledge graph of the industrial equipment; s2.2, carrying out community division on the industrial equipment operation and maintenance knowledge graph by adopting a graph community discovery algorithm Leiden according to the relation between entity nodes to obtain a plurality of mutually-distinguished knowledge communities, and using Represent the first The community of knowledge of the individual community of knowledge, Representing the total number of knowledge communities.
4. The intelligent query and answer method for equipment operation and maintenance based on knowledge-graph context fusion according to claim 1, wherein the specific implementation process of the step S3 is as follows: s3.1, receiving an industrial equipment operation and maintenance problem input by a user, encoding the industrial equipment operation and maintenance problem by adopting a text embedding model, and generating a problem vector for representing semantic features of the problem ; S3.2, extracting semantic attribute features and topological structure features in the knowledge communities aiming at each knowledge community, and constructing initial graph feature representation of the knowledge communities, wherein the semantic attribute features comprise node feature matrixes generated based on entity descriptions in the communities, and the topological structure features comprise adjacent information for representing connection relations among entities in the communities; S3.3, processing the initial graph characteristic representation by utilizing a graph neural network, gathering neighborhood characteristics through message transmission among nodes, capturing topology semantic association in communities, performing global pooling on the gathered characteristics, generating community vectors representing the overall characteristics of knowledge communities, and using Represent the first Community vectors for individual knowledge communities.
5. The intelligent query and answer method for equipment operation and maintenance based on knowledge-graph context fusion according to claim 4, wherein the specific implementation process of step S4 is as follows: s4.1, setting a differentiated similarity measurement measure set Wherein The K-th similarity measure is represented, K represents the number of the used similarity measure, and the correlation between the problem vector and the knowledge community vector is described in different similarity measures from different angles, wherein cosine similarity is used for measuring consistency of semantic directions, european similarity is used for describing the overall semantic position proximity degree, manhattan similarity is used for reflecting the accumulation condition of multi-dimensional semantic feature differences, dot product similarity is used for representing the activation strength of the problem semantics on the community semantic features, and a multi-angle and multi-layer correlation result is formed; S4.2 for the kth similarity metric measure Respectively calculating problem vectors Vector of community with knowledge Similarity between them by Representing problem vectors and the first Obtaining similarity results of the k-th measure and the community similarity result set Wherein Representing calculation of problem vectors with kth measure And the first Community vector for individual knowledge communities Similarity between; S4.3 similarity results set for kth measure Selecting the top of similarity ranking under the kth measure according to the similarity from large to small Sequentially acquiring candidate community sets of all measures, and then taking union sets for all candidate community sets to form a joint candidate community set I.e. identifying the frame, generating by means of combination enumeration Is composed of all mathematical subsets of Power set of , Is a collection The number of elements in the matrix; S4.4, normalizing the similarity result in the candidate community set corresponding to the kth measure to obtain similarity reliability distribution of the kth measure definition on the combined candidate community set And for the knowledge communities belonging to the k measure candidate community set, the confidence value is determined by the normalized similarity result, and for the knowledge communities belonging to the joint candidate community set but not to the k measure candidate community set, the confidence value is set to 0, and the similarity confidence distribution of all measures is sequentially acquired.
6. The intelligent query and answer method for equipment operation and maintenance based on knowledge-graph context fusion according to claim 5, wherein the specific implementation process of step S5 is as follows: S5.1, for the kth measure, determining discount factors for each knowledge community in the joint candidate community set respectively based on the distribution form index and combined with the confidence value of each joint candidate community in the similarity confidence distribution corresponding to the kth measure to form a discount parameter vector under the kth measure Wherein Indexing the variables for positive integers, representing knowledge community numbers in the joint candidate community set, Represent under the kth measure Discount factors of the corresponding credibility of the knowledge communities; s5.2 for the kth measure, according to the discount parameter vector For similarity confidence distribution The corresponding credibility value of each knowledge community is discounted item by item to obtain the credibility distribution of the community credibility discounted The discount process is implemented by a set transformation operator for any subset The confidence value after community confidence discount is composed of all subsets The original credibility value is obtained by weighting and accumulating, wherein For a pair of Is weighted by discount kernel function Control of discount kernel function according to And (3) with The collection contains a relation, and product combination is carried out on discount factors of the related knowledge community corresponding credibility and retention coefficients of the knowledge community corresponding credibility, so that the weighted propagation of community-level credibility in a collection space is realized; S5.3, outputting K similarity reliability distributions subjected to community reliability discount, and taking the K similarity reliability distributions as input for fusion of a plurality of subsequent similarity reliability distributions.
7. The intelligent query and answer method for equipment operation and maintenance based on knowledge-graph context fusion according to claim 6, wherein the specific implementation process of step S5.1 is as follows: selecting a knowledge community with a confidence value greater than 0 in the similarity confidence distribution as an effective community, calculating information entropy of the confidence distribution of the effective community according to the duty ratio of the confidence value of the effective community in the total confidence of the effective community, and obtaining a discrete degree index through maximum entropy normalization processing Sequencing the effective community credibility from large to small, and calculating the difference between the maximum credibility and the next-largest credibility to obtain a distinguishing strength index ; According to the discrete degree index Differentiating intensity index Will be And (3) with Weighted summation is carried out to construct an interval adjustment coefficient corresponding to the kth measure Preset discount zone Defining the desirable range of the discount factor in the interval Determining discount interval under kth measure according to interval adjustment coefficient Wherein Equal to , Equal to And (3) with Is a summation result of (a); In the discount zone of the kth measure In the method, linear mapping is carried out according to the credibility value of each knowledge community under the kth measure to obtain discount factors corresponding to each knowledge community, and the knowledge communities under the kth measure Corresponding discount factors Equal to And (3) with Is used to sum the results of the (c), Representing knowledge communities in a set of joint candidate communities under a kth measure The corresponding confidence value is used to determine the confidence level, Indexing a variable for a positive integer for representing knowledge community numbers in the joint candidate community set; obtaining discount parameter vector under kth measure , Represent under the kth measure The individual knowledge communities correspond to discounting factors for confidence.
8. The intelligent query and answer method for equipment operation and maintenance based on knowledge-graph context fusion according to claim 7, wherein the specific implementation process of step S6 is as follows: S6.1, fusing K similarity reliability distributions output by S5 after community reliability discount by using a combination rule in a D-S evidence theory to obtain comprehensive reliability distribution; s6.2, quantizing the comprehensive reliability distribution through an approximate probability function to obtain quantized comprehensive reliability distribution, and performing a fuzzy algorithm on the comprehensive reliability distribution Sequencing the confidence results of single point subsets in the table, and taking the confidence results before ranking Based on the target knowledge community, extracting knowledge content related to the problem from the corresponding industrial equipment operation and maintenance knowledge graph, and constructing a knowledge graph context for inputting the industrial equipment operation and maintenance large language model; And S6.3, carrying out parameter fine adjustment on the large language model by using the industrial equipment operation and maintenance question-answering sample obtained in the S1 to obtain the industrial equipment operation and maintenance large language model, inputting the knowledge graph context obtained in the S6.2 and the problem input by the user into the industrial equipment operation and maintenance large language model, and generating natural language answers corresponding to the industrial equipment operation and maintenance problem after the large model is inferred.
9. The intelligent equipment operation and maintenance question-answering system based on knowledge graph context fusion is used for realizing the intelligent equipment operation and maintenance question-answering method of any one of claims 1 to 8, and is characterized by comprising the following modules: The knowledge text module is used for constructing a knowledge text data set in the field of operation and maintenance of the industrial equipment; the knowledge community module is used for constructing an operation and maintenance knowledge graph of the industrial equipment based on the knowledge text data set, and carrying out community division on the knowledge graph to obtain a plurality of knowledge communities representing knowledge units with different operation and maintenance; the problem vector and community vector module is used for vectorizing the problem input by the user and the knowledge community respectively to obtain a problem vector and a community vector; The similarity credibility distribution module is used for calculating the diversity association relation between the problem vector and the community vector through the differentiated similarity measurement and generating similarity credibility distribution reflecting different association degrees; The adaptability adjustment module is used for adaptively adjusting the similarity reliability distribution through a community reliability discount strategy; And the industrial equipment operation and maintenance question answering module is used for fusing the similarity reliability distribution subjected to the adaptive adjustment, constructing a knowledge graph context, inputting the knowledge graph context into an industrial equipment operation and maintenance large language model, and generating an answer of the industrial equipment operation and maintenance question.

Description

Intelligent device operation and maintenance question-answering method and system based on knowledge graph context fusion Technical Field The invention relates to the technical field of artificial intelligence and natural language processing, in particular to a device operation and maintenance intelligent question-answering method and system based on knowledge graph context fusion. Background The large industrial equipment is widely applied to the key industrial fields of chemical industry, energy sources, electric power, traffic and the like, such as a wind generating set, a gas turbine, a large compressor and the like, and generally operates in production scenes with complex environments and changeable working conditions, and the operation state of the large industrial equipment is directly related to production safety and economic benefits. Once equipment is in fault or abnormal state, operation and maintenance personnel are often required to complete fault positioning, cause analysis and treatment scheme formulation in a short time, and high requirements are put on operation and maintenance efficiency and professional experience. In the operation and maintenance process of the existing industrial equipment, related knowledge is mainly stored in unstructured or weakly structured texts such as technical manuals, operation and maintenance rules, historical maintenance records and the like in a scattered manner, operation and maintenance personnel usually rely on manual retrieval or expert experience to conduct problem treatment, and quick and accurate response to complex problems is difficult to achieve. In recent years, large language models have been introduced into industrial operation and maintenance scenarios to assist personnel in problem consultation and solution generation due to their prominence in natural language understanding and knowledge questions and answers. However, the existing operation and maintenance question-answering mode based on the large language model generally has the problems of insufficient knowledge dependence on the professional field, difficult assessment of knowledge source credibility, insufficient stability of reasoning results, even illusion and the like, and restricts the application effect of the operation and maintenance question-answering mode in the actual industrial equipment operation and maintenance scene. Disclosure of Invention In order to solve the problems of dispersion of equipment operation and maintenance knowledge, single problem matching, insufficient reliability of reasoning results and the like in the prior art, the invention provides a method and a system for intelligent question-answering of equipment operation and maintenance based on knowledge graph context fusion, according to the method, the user problem is matched with the knowledge community in a multidimensional mode through the differentiated similarity measurement, the reliability of the constructed knowledge graph context is improved by combining community credibility discount and credibility fusion strategy to conduct credibility regulation on the search result, and support is provided for equipment operation and maintenance decision and intelligent question-answering. In one aspect of the invention, an intelligent device operation and maintenance question-answering method based on knowledge graph context fusion is provided, and the method comprises the following steps: S1, constructing a knowledge text data set in the field of operation and maintenance of industrial equipment. And S2, constructing an industrial equipment operation and maintenance knowledge graph based on the knowledge text data set obtained in the step S1, and carrying out community division on the knowledge graph to obtain a plurality of knowledge communities representing knowledge units with different operation and maintenance. And S3, respectively carrying out vectorization representation on the problem input by the user and the knowledge community obtained in the step S2 to obtain a problem vector and a community vector. And S4, calculating the diversity association relation between the problem vector and the community vector obtained in the S3 through the differentiated similarity measurement, and generating similarity reliability distribution reflecting different association degrees. And S5, providing a community credibility discount strategy, and adaptively adjusting the similarity credibility distribution obtained in the step S4 so as to improve the reliability of the credibility distribution. And S6, fusing the similarity reliability distribution subjected to the adaptive adjustment, constructing a knowledge graph context according to the fusion result, and inputting the obtained knowledge graph context into the large language model of the operation and maintenance of the industrial equipment so as to generate answers of the operation and maintenance questions of the industrial equipment. Wherein, the step S3 sp