CN-122022554-A - QKK product index association analysis method, system and storage medium based on semantic recognition

CN122022554ACN 122022554 ACN122022554 ACN 122022554ACN-122022554-A

Abstract

The invention provides a QKK product index association analysis method, a system and a storage medium based on semantic recognition, which relate to the technical field of communication network operation and maintenance and comprise the steps of collecting QKK index data from a plurality of heterogeneous data sources for association alignment; extracting numerical semantic vectors and text semantic vectors, mapping the numerical semantic vectors and the text semantic vectors to a unified semantic vector space through a cross-modal semantic alignment network, constructing an initial ontology based on field standards, finding new concepts and new relation candidates through clustering processing and a graph neural network, updating the new concepts and the new relation candidates to a dynamic semantic ontology library after man-machine collaborative verification, constructing a semantic association graph through ontology instantiation, performing graph traversal on a target event to obtain an association path, and outputting a root cause analysis result by combining a causal discovery algorithm with multi-evidence fusion. The invention solves the problem of heterogeneous data semantic gap, realizes the dynamic evolution of the ontology and root cause positioning at the causal level, and remarkably improves the root cause analysis accuracy.

Inventors

WANG PENGLIANG
CHEN XI
ZHANG LEI
YANG CHAO
ZHANG JING

Assignees

广州丰石科技有限公司

Dates

Publication Date: 20260512
Application Date: 20260106

Claims (10)

1. A QKK product index association analysis method based on semantic recognition is characterized by comprising the following steps: s1, acquiring QKK index data from a plurality of heterogeneous data sources, cleaning and fusing the acquired data, and carrying out association alignment on the data from different sources by taking a timestamp and a network element identifier as key keys; s2, carrying out feature extraction on a numerical sequence index by adopting a time sequence coding model to obtain a numerical semantic vector, carrying out named entity recognition on text data by adopting an entity recognition model, and obtaining a text semantic vector by adopting the text coding model; S3, constructing an initial ontology based on field standards and expert experience, defining concept classes and basic relations, clustering vectors in the unified semantic vector space to find new concept candidates, analyzing an ontology graph structure by using a graph neural network to find new relation candidates, confirming the new concept candidates and the new relation candidates through man-machine collaborative verification, and updating the confirmed concept and relation to a dynamic semantic ontology library; And S4, instantiating concepts and relations in the dynamic semantic ontology library to construct a semantic association graph, traversing the graph in the semantic association graph by taking a node corresponding to a target event as a starting point when the target event occurs to acquire an association path, performing causal reasoning on the association path by applying a causal discovery algorithm, sequencing candidate root causes by combining multi-evidence fusion, and outputting a root cause analysis result.
2. The QKK product index association analysis method based on semantic recognition according to claim 1 is characterized in that in S2, the time sequence coding model is TimesNet model or a transducer-based encoder, the entity recognition model is MSFM model which is pre-trained and enhanced by corpus in the telecommunication field, and the text coding model is Sentence-BERT model.
3. The QKK product index association analysis method based on semantic recognition according to claim 1, wherein in S2, the cross-modal semantic alignment network is trained in a contrast learning mode, and the training aims at maximizing vector similarity of matched pairs and minimizing vector similarity of unmatched pairs, wherein the matched pairs comprise paired numerical sequences and descriptive texts.
4. The QKK product index association analysis method based on semantic recognition according to claim 1, wherein in the step S3, a DBSCAN density clustering algorithm is adopted for clustering, when a newly formed cluster meets a compactness condition and the distance between the newly formed cluster and the center of an existing concept cluster exceeds a preset threshold, the corresponding cluster is marked as a new concept candidate, and the graph neural network learns propagation and association modes among nodes in an ontology graph, and suggests to add new relationship edges among specific nodes or adjust confidence weights of the existing relationships.
5. The QKK product index association analysis method based on semantic recognition as set forth in claim 1, wherein the step S3 further comprises a closed loop optimization step of reflowing feedback signals of experts in a human-computer collaborative verification process to a model optimizer for optimizing parameters of the clustering process and the graph neural network.
6. The QKK product index association analysis method based on semantic recognition according to claim 1, wherein in the S4, the causal discovery algorithm is a PC algorithm, whether direct causal relationship exists among variables is judged through condition independence test, and the multi-evidence fusion comprises ranking candidate root causes according to comprehensive causal strength, associated edge weight, network topology distance and historical occurrence frequency.
7. The QKK product index association analysis method based on semantic recognition according to claim 1, further comprising S5, presenting the root cause analysis result through a visual interface, receiving evaluation feedback of an operation and maintenance expert on the root cause analysis result, and returning the evaluation feedback to the dynamic semantic ontology library and the causal reasoning process for optimization.
8. A QKK product index association analysis system based on semantic recognition for implementing the analysis method of any one of claims 1 to 7, comprising: The data acquisition and preprocessing module is configured to acquire QKK index data from a plurality of heterogeneous data sources, carry out cleaning and fusion processing on the acquired data, and carry out association alignment on the data of different sources by taking the time stamp and the network element identifier as key keys; The multi-modal semantic identification module is configured to extract the characteristics of the numerical sequence index by adopting a time sequence coding model to obtain a numerical semantic vector, carrying out named entity identification on text data by adopting an entity identification model, obtaining a text semantic vector through the text coding model, and mapping the numerical semantic vector and the text semantic vector into a unified semantic vector space through a cross-modal semantic alignment network; The dynamic ontology construction module is configured to construct an initial ontology based on field standards and expert experience, define concept classes and basic relations, perform clustering processing on vectors in the unified semantic vector space to find new concept candidates, analyze an ontology graph structure by using a graph neural network to find new relation candidates, and confirm and update the new concept candidates and the new relation candidates to a dynamic semantic ontology library through man-machine collaborative verification; And the association analysis and causal reasoning module is configured to instantiate concepts and relations in the dynamic semantic ontology library to construct a semantic association graph, when a target event occurs, traversing the graph by taking a node corresponding to the target event as a starting point to obtain an association path, performing causal reasoning on the association path by applying a causal discovery algorithm, and sequencing candidate root causes by combining multi-evidence fusion to output root cause analysis results.
9. The QKK product index association analysis system based on semantic recognition of claim 8 further comprising a visualization and feedback module configured to visually present the root cause analysis results, receive evaluation feedback from an operation and maintenance expert, and return to the dynamic ontology construction module and the association analysis and causal reasoning module for optimization.
10. A computer readable storage medium, characterized in that the computer readable storage medium has stored thereon a computer program which, when executed by a processor, implements the method of any of claims 1-7.

Description

QKK product index association analysis method, system and storage medium based on semantic recognition Technical Field The invention relates to the technical field of communication network operation and maintenance, in particular to a QKK product index association analysis method, a system and a storage medium based on semantic recognition. Background The telecommunications network index management technology has undergone three stages of evolution. The early stage mainly depends on personal experience of a senior engineer, acquires indexes through an SNMP protocol and sets a fixed threshold alarm, and the correlation analysis completely depends on manual logic inference, so that the problems of low efficiency, poor consistency, incapability of precipitating knowledge and the like exist. Then, entering a statistical analysis stage, an operator starts to find the linear relation among indexes by using statistical methods such as pearson correlation coefficient, principal component analysis and the like, and uses ARIMA, prophet isochronous models to conduct trend prediction, but the correlation is not equal to the causal relation, and unstructured data is difficult to process. At present, in an intelligent operation and maintenance germination stage, machine learning models such as isolated forests, LSTM and the like are introduced into the operation and maintenance field, but challenges such as model fragmentation, semantic understanding deficiency, poor interpretability and the like are faced, conclusions obtained by different models can conflict with each other, and the models cannot understand the logical relationship between service indexes and user experience. Current telecom operators face four levels of core pain points in QKK index association analysis. In the data layer, QKK index sources are wide, equipment-level KPIs of a network management system, service-level KPIs of a service platform, qoE data of a user experience system, work order logs of a supporting system and the like are covered, the data are different in format and sampling frequency, naming standards and service connotations are greatly different, and the prior art lacks a unified means to understand and align service semantics behind heterogeneous data. At the model level, the existing system mostly relies on a predefined association rule base, the rules are written by experts according to historical experience, the network needs to be re-examined and updated by the experts every time the network is upgraded or a new service is on line, the response speed is low, all fault scenes cannot be covered, and the rule base cannot be automatically learned from new data. In the analysis level, most of the existing tools only answer whether the indexes are related or not, but cannot judge whether the causal relationship exists or not, and the confusion of the relativity and the causal relationship can lead to inaccurate root cause positioning and incorrect operation and maintenance action direction. In the knowledge level, expert knowledge accumulated in the operation and maintenance field is difficult to effectively inject into a data driving model, the result is uncontrollable due to complete dependence on data driving, or the system is stiff due to complete dependence on rules, and knowledge guidance and data driving are difficult to deeply fuse. The Chinese patent document CN119961409A discloses an intelligent operation and maintenance optimization method based on a semantic enhancement knowledge graph, and discloses a technical scheme of extracting text key entities through named entity recognition, constructing an ontology model description operation and maintenance field concept system, analyzing operation and maintenance personnel input by using an NLU model and carrying out reasoning judgment based on the knowledge graph, which has the technical effects of accurately analyzing problem description and command instructions, avoiding complex sentence structure misidentification and supporting data periodic update, but still has the problems that only text type single-mode data can be processed, numerical time sequence indexes can not be fused, the ontology model lacks dynamic evolution capability depending on manual maintenance expansion, and a reasoning mechanism can not distinguish correlation and causality based on rule link tracking. The Chinese patent document CN120632061B discloses an intelligent analysis method and system for enterprise digital transformation, and discloses a technical scheme which is characterized in that scene factors are grasped through multi-channel data, knowledge maps are constructed, deep learning semantic feature extraction is utilized for carrying out association matching, rule reasoning and graph structure reasoning model analysis are adopted for direct and indirect influence, the technical effects of carrying out deep association matching on the scene factors and the enterprise knowledge, generating acc