Self-planning-self-verification network operation and maintenance knowledge graph generation method based on thinking chain

Abstract

The invention relates to the technical field of artificial intelligence and network operation and maintenance, in particular to a thinking chain-based self-planning-self-verification network operation and maintenance knowledge graph generation method, which comprises the following steps of: constructing an initial knowledge graph as a reference by taking a small amount of high-quality expert annotated question and answer pairs and a tool table as starting points; structural samples are synthesized in a large scale through an analysis-planning-execution thinking chain self-planning process driven by a large language model, then two-stage self-verification of screening based on similarity distribution and correction based on content verification is carried out on synthesized data, and finally the verified high-quality samples are mapped into triples to be merged into an initial graph; therefore, a closed loop of'initial graph construction-data synthesis-knowledge verification-incremental graph entering 'is formed, controllability of graph expansion and correctness of professional operation and maintenance knowledge are guaranteed, high-quality automatic expansion of the network operation and maintenance knowledge graph is realized under the condition of low labeling cost, and reliable knowledge support is provided for intelligent diagnosis and automatic operation and maintenance of operators.

Description

Technical Field

[0001] This invention relates to the fields of artificial intelligence and network operation and maintenance technology, specifically to a method for generating a self-planning and self-verifying network operation and maintenance knowledge graph based on thought chains. Background Technology

[0002] As the scale and complexity of operator communication networks continue to rise, the reliance on automated diagnosis and intelligent decision-making in network operation and maintenance is becoming increasingly significant. To support such applications, knowledge graphs, as a structured representation, have become a common means of building network fault knowledge bases and enhancing operator-specific question-and-answer and diagnostic systems because they can explicitly describe entities and relationships and carry reasoning paths. In terms of knowledge graph expansion and synthetic data generation, there are currently two main methods: synthetic data generation and bootstrapping methods based on large language models. These methods generate instruction samples through model bootstrapping to expand the training set size and alleviate the problem of scarce annotations. At the same time, they guide large language models to generate structured or semi-structured text samples through prompt design and few-sample examples for downstream training or annotation assistance, such as the Self-Instruct method.

[0003] Knowledge graph-driven generation and retrieval enhancement methods utilize existing graphs or graph structures to guide sampling and generation, making the output more multi-hop knowledge coverage and semantic coherence. Retrieval enhancement generation concatenates the context retrieved from external documents or graph structures with a large language model before inputting it to improve the accuracy and factuality of generation, such as the GraphGen method.

[0004] However, when the above methods are directly applied to the expansion of network operation and maintenance knowledge graphs in engineering, they still face the following problems: the cost of seed data and annotation is relatively large, high-quality operation and maintenance knowledge relies more on manual annotation by domain experts, and although self-instruction and other bootstrap data synthesis methods can expand the sample size, the quality of the generated samples is uneven, and a lot of complex screening and cleaning are still required, making it difficult to fundamentally reduce the cost of annotation.

[0005] The logical consistency with facts in synthetic data is insufficient. Although automatically synthesized data based on large language models has advantages in semantic coverage, it is prone to problems such as fictitious tools, inference contradictions, or inconsistent parameter references. Existing filtering or simple retrieval methods cannot fully guarantee the correctness of synthetic data in terms of logical rigor and tool parameter dependencies. To address this, a self-planning and self-verifying network operation and maintenance knowledge graph generation method based on thought chains is proposed. Summary of the Invention

[0006] To address the technical problems existing in the prior art, this invention provides a method for generating a self-planning and self-verifying network operation and maintenance knowledge graph based on thought chains.

[0007] To address the aforementioned technical problems, this invention provides the following technical solution: a self-planning and self-verifying network operation and maintenance knowledge graph generation method based on thought chains, wherein the network operation and maintenance knowledge graph generation method specifically includes the following steps:

[0008] S1. Gather business data from multiple scenarios, define the ontology information of the knowledge graph, and construct the initial knowledge graph;

[0009] S2: First, perform structural analysis on high-quality seed question-answer pairs and tool tables. After analysis, perform synthesis planning to generate a synthesis plan. Drive the big oracle model to generate synthetic data according to the synthesis plan.

[0010] S3 selects data at the form and distribution levels for the synthesized data, and self-verifies and corrects the synthesized data at the content and structure levels to obtain high-quality synthesized data.

[0011] S4 uses regular expressions to parse the ontology information of the knowledge graph in the high-quality synthesized data string. The ontology information includes triples of typical questions, scenarios, sub-scenarios, thought chain steps, and tools. The triples are incrementally added to the initial knowledge graph according to the mapping rules to obtain the updated knowledge graph.

[0012] Preferably, step S1 specifically includes the following steps:

[0013] S11 consists of multi-scenario business data information compiled by business experts in the field of network operation and maintenance technology, and the ontology information of the knowledge graph jointly defined by algorithm experts.

[0014] The multi-scenario business materials specifically include business process diagrams, process tables, typical operation and maintenance questions and standardized answer templates;

[0015] S12, organize the seed question-and-answer pairs data annotated by business experts and the ontology information used to build the knowledge graph;

[0016] S13. After sorting, each seed question-answer pair is parsed using regular expressions to extract specific information about the knowledge graph ontology from the string. Then, based on the extracted specific data, initial entity-relation triples are generated to construct the initial knowledge graph.

[0017] Preferably, step S2 specifically includes the following steps:

[0018] S21, Structural Analysis: Input high-quality seed question-answer pairs and tool tables labeled by several business experts into the large language model. Guide the large language model to extract question patterns, element co-occurrence relationships, typical reasoning paths and tool call paradigms from the high-quality seed question-answer pairs and tool tables through the templated Prompt of the large language model, and thereby construct a reusable structural element library.

[0019] The structural element library includes template definitions, parameter dependency mappings, and templates for commonly used tools.

[0020] S22, Synthesis Planning: Based on seed data samples, structural element library and prompt word templates, the Prompt of the large language model is constructed. It provides seed data and drives the large language model to construct an executable synthesis plan based on the seed data and synthesis data.

[0021] The specific requirements for the synthesis plan specified in the Prompt are: adding sub-scene descriptions, semantic variant strategies, element replacement rules, slot definitions and replacement lists, and an upper limit on the number of samples generated for each type, in order to control the generation strategy and prevent pattern collapse;

[0022] S23, Perform synthesis: The large language model’s Prompt drives the large language model to generate synthesized data in batches according to the seed data samples and synthesis plan. The output of synthesized data includes: structured knowledge graph triple information including typical questions, scenarios, sub-scenarios, thought chain steps and tools.

[0023] During execution, multiple sets of simple random sampling with replacement are performed on the high-quality seed question-and-answer dataset to generate seed data samples. Based on the required number of data to be generated, an upper limit is set for the number of high-quality seed questions and answers generated, thereby improving sample diversity and reducing the risk of homogenization.

[0024] Preferably, in step S3, when selecting high-quality data at the form and distribution levels, a high-quality data selector is used to filter out synthetic data that is close to the semantic feature distribution of the seed data, that is, to select high-quality synthetic data.

[0025] At runtime, regular expression matching is used to parse the string "sub-scene - tool" from each synthesized data text. This string is then concatenated into a comparison text for further analysis. A text encoding model is used to calculate the text embedding vector, mapping the comparison text to a vector space. During computation, the embedding vector for each synthesized data text is calculated. Similarity sequence with all seed samples Specifically:

[0026]

[0027] in, Represents synthetic data Similarity with the nth seed sample;

[0028] The similarity scores of all synthetic data and seed samples are combined to form a global sample similarity sequence. Then, the distribution of the global sample similarity sequence is fitted with a normal distribution to obtain the overall distribution characteristics of the global sample similarity sequence.

[0029] The Spearman correlation coefficient is used to quantify the consistency between the distribution of a single sample similarity sequence and the distribution of the entire set of sample similarity sequences. As an indicator of consistency, the retention threshold derived from the population statistic is specifically:

[0030]

[0031] in, and Representing all samples The mean and standard deviation, This indicates an adjustable empirical coefficient;

[0032] Samples that satisfy this inequality are retained as initial screening samples, while synthetic data that significantly deviates from the distribution of the seed samples, is semantically irrelevant, or is repetitive are statistically removed.

[0033] Preferably, in step S3, when performing self-verification and correction on the synthetic data at the content and structure levels, the synthetic data self-verifier sequentially performs rule-based repair, content quality assessment, and low-score automatic correction on the synthetic data.

[0034] When the synthetic data self-verifier runs, it first performs consistency verification on all synthetic data structures and performs rule-based repair: when performing rule-based repair, it first locates samples that are inconsistent with the seed data format or have inconsistent parameter references.

[0035] After localization, the system automatically performs initial screening on samples that are inconsistent with the seed data format and have inconsistent parameter references by calling the rule repair script, thus achieving structural repair without the need for a large language model.

[0036] After fixing the format and parameter issues, business experts led the design of review rules and standards tailored to the operation and maintenance scenarios. A high-performance large language model was invoked to automatically perform content quality assessments according to a templated prompt. The assessments were conducted from four operation and maintenance-specific dimensions: the rigor of the handling logic, the coverage of the fault analysis, the compliance of the operation and maintenance report, and the fit of the operation and maintenance knowledge base. A five-point quantitative scoring system was used to generate specific scoring opinions. During the scoring process, the review input included the sample to be verified and several most similar sub-scenario graphs retrieved from the initial graph. When the total score of the sample was lower than the set threshold, an automated correction process for low-scoring samples was triggered. The large language model combined the sample data, scoring opinions, and context to perform targeted revisions and output the corrected samples.

[0037] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0038] 1. This invention utilizes the reasoning capabilities of a large language model to drive the generation process of the "analysis-planning-execution" thought chain through structured prompt words. It generates diverse operation and maintenance data from a small number of seed questions and answers, and ensures the consistency of the structure and semantic diversity of the diverse operation and maintenance data, effectively alleviating the problem of insufficient coverage caused by the scarcity of domain annotations.

[0039] 2. This invention employs a dual verification method consisting of similarity distribution screening and content verification and automatic correction designed by operation and maintenance business experts to automatically verify and repair generated samples at the logical and knowledge levels, ensuring data consistency and accuracy. It can automatically identify and repair low-quality samples from both statistical distribution and content quality dimensions, significantly reducing the risk of common illusions and inference contradictions in large language model generated data, and achieving knowledge self-verification and self-correction.

[0040] 3. This invention structures verified high-quality data into triples and incorporates them into the knowledge graph, enabling continuous incremental updates and dynamic evolution of knowledge. It achieves high-quality automatic expansion of the network operation and maintenance knowledge graph under low annotation costs, providing reliable knowledge support for operators' intelligent diagnosis and automated operation and maintenance. Attached Figure Description

[0041] Figure 1 This is a schematic diagram of the self-planning-self-verifying network operation and maintenance graph generation architecture based on the thought chain of the present invention.

[0042] Figure 2 This is a schematic diagram of the Spearman correlation coefficient distribution of the synthetic data of this invention;

[0043] Figure 3 This is a scoring table for the synthetic data before correction in this invention;

[0044] Figure 4 This is a scoring table for the corrected synthetic data of the present invention. Detailed Implementation

[0045] The present invention will be further described below with reference to the accompanying drawings and embodiments, which illustrate the above and other technical features and advantages of the present invention. However, the following embodiments are merely preferred embodiments of the present invention and are not exhaustive.

[0046] Example 1:

[0047] like Figure 1 As shown, this invention provides a method for generating a self-planning and self-verifying network operation and maintenance knowledge graph based on thought chains. The method for generating a network operation and maintenance knowledge graph specifically includes the following steps:

[0048] S1. Gather business data from multiple scenarios, define the ontology information of the knowledge graph, and construct the initial knowledge graph;

[0049] S2: First, perform structural analysis on high-quality seed question-answer pairs and tool tables. After analysis, perform synthesis planning to generate a synthesis plan. Drive the big oracle model to generate synthetic data according to the synthesis plan.

[0050] S3 uses a high-quality data selector to select data at the form and distribution levels of the synthetic data, and uses a synthetic data verifier to self-verify and correct the synthetic data at the content and structure levels, thus obtaining high-quality synthetic data.

[0051] S4. After the high-quality data selector selects and the synthetic data sub-verifier performs self-verification and correction, the high-quality synthetic data is parsed using regular expressions to extract the ontology information of the knowledge graph from the string. The ontology information includes triples of typical questions, scenarios, sub-scenarios, thought chain steps, and tools. The triples are incrementally added to the initial knowledge graph according to the mapping rules to obtain the updated knowledge graph.

[0052] Application scenarios include fiber optic cable breakage location and network health analysis; sub-scenarios are more specific problem types within a scenario, such as querying the status of a specific device or querying alarms; the thought chain steps are the methods and steps for solving problems in sub-scenarios, and tools include time calibration tools and network element alarm query tools.

[0053] In this embodiment, step S1 specifically includes the following steps:

[0054] S11 consists of multi-scenario business data information compiled by business experts in the field of network operation and maintenance technology, combined with ontology information of the knowledge graph jointly defined by algorithm experts;

[0055] The multi-scenario business materials specifically include business process diagrams, process tables, typical operation and maintenance questions and standardized answer templates;

[0056] S12, organize the seed question-and-answer pairs data annotated by business experts and the ontology information used to build the knowledge graph;

[0057] S13. After sorting, each seed question-answer pair is parsed using regular expressions to extract specific information of the knowledge graph ontology from the string (this ontology contains five types of nodes: typical questions, scenarios, sub-scenarios, thought chain steps, and tools). Then, based on the extracted specific data, initial entity-relationship triples are generated to construct the initial knowledge graph.

[0058] In this embodiment, step S2 specifically includes the following steps:

[0059] S21, Structural Analysis: Input high-quality seed question-answer pairs and tool tables labeled by several business experts into the large language model. Guide the large language model to extract question patterns, element co-occurrence relationships, typical reasoning paths and tool call paradigms from the high-quality seed question-answer pairs and tool tables through the templated Prompt of the large language model, and thereby construct a reusable structural element library.

[0060] The structural element library includes template definitions, parameter dependency mappings, and templates for commonly used tools.

[0061] S22, Synthesis Planning: Based on seed data samples, structural element library and prompt word template (i.e., a general prompt word template for a large language model containing identity definition + core task + constraints + output format structure), construct the Prompt of the large language model, provide seed data, and drive the large language model to generate an executable synthesis plan based on the seed data and Prompt;

[0062] The specific requirements for the synthesis plan outlined in the Prompt are: adding sub-scene descriptions, semantic variant strategies, element replacement rules, slot definitions and replacement lists, and an upper limit on the number of samples generated for each type, in order to control the generation strategy and prevent pattern collapse;

[0063] S23, Perform synthesis: The large language model’s Prompt drives the large language model to generate synthesized data in batches according to the seed data samples and synthesis plan. The output of synthesized data includes: structured knowledge graph triple information including typical questions, scenarios, sub-scenarios, thought chain steps and tools.

[0064] During execution, multiple sets of simple random sampling with replacement are performed on the high-quality seed question-and-answer dataset to generate seed data samples. Based on the required number of data to be generated, an upper limit is set for the number of high-quality seed questions and answers to be generated, such as 5 times. Each seed data sample in the sampling pool will have 5 samples, which is simple random sampling without replacement, thereby improving sample diversity and reducing the risk of homogenization.

[0065] When performing simple random sampling with replacement, a seed data sampling pool is first constructed using a Python script; then, the built-in random function of Python is called to randomly select two samples from the sampling pool. After selection, the two samples are not removed, and they are returned to the seed data sampling pool.

[0066] A Prompt is an instruction, question, or instruction given to the Big Prophet model. The Big Prophet model outputs corresponding content based on the input signals given by the Prompt. A templated Prompt is a Prompt that conforms to the general structure template of prompt words in recent big language models. A templated Prompt includes identity definition, core tasks, constraints, and output format.

[0067] In this embodiment, in step S3, when selecting high-quality data at the form and distribution levels, a high-quality data selector filters out synthetic data that is close to the semantic feature distribution of the seed data, that is, high-quality synthetic data is selected.

[0068] At runtime, regular expression matching is used to parse the string "sub-scene – tool" (part of the triplet ontology) from each synthesized data text. This string is then concatenated into a comparison text for contrast. A text encoding model is used to calculate the text embedding vector, mapping the comparison text to a vector space. During computation, the embedding vector for each synthesized data text is calculated. Similarity sequence with all seed samples Specifically:

[0069]

[0070] in, Represents synthetic data Similarity with the nth seed sample;

[0071] The similarity scores of all synthetic data and seed samples are merged to form a global sample similarity sequence. Then, a normal distribution is fitted to the distribution of this global sample similarity sequence to obtain its overall distribution characteristics, specifically: mean... with standard deviation ;

[0072] The Spearman correlation coefficient is used to quantify the consistency between the distribution of a single sample similarity sequence and the distribution of the entire set of sample similarity sequences. As an indicator of sample distribution consistency, it measures The degree of correlation with the overall similarity ranking, i.e., the retention threshold derived from the overall statistics, is as follows:

[0073]

[0074] in, and Representing all samples The mean and standard deviation, This indicates an adjustable empirical coefficient;

[0075] Samples that satisfy this inequality are retained as initial screening samples, while synthetic data that significantly deviates from the distribution of the seed samples, is semantically irrelevant, or is repetitive are statistically removed.

[0076] In this embodiment, in step S3, when performing self-verification and correction on the synthetic data at the content and structure levels, the synthetic data self-verifier sequentially performs rule-based repair, content quality assessment, and low-score automatic correction on the synthetic data.

[0077] When the synthetic data self-verifier runs, it first performs consistency verification on all synthetic data structures and performs rule-based repair: when performing rule-based repair, it first locates samples that are inconsistent with the seed data format or have inconsistent parameter references.

[0078] After localization, the system automatically performs initial screening on samples that are inconsistent with the seed data format and have inconsistent parameter references by calling the rule repair script, thus achieving structural repair without the need for a large language model.

[0079] The rule repair script corrects the format of the synthesized data to be consistent with the format of the seed data. Only after the format of the synthesized data and the seed sample data are unified can the information be extracted by string regular expression matching and the knowledge graph ontology triple information in the synthesized data be parsed.

[0080] After format and parameter repair, business experts lead the design of review rules and standards tailored to the operation and maintenance scenario. A high-performance large language model is invoked to automatically perform content quality assessment according to a templated prompt. The assessment evaluates content from four operation-specific dimensions: the rigor of the handling logic, the coverage of fault analysis, the compliance of the operation and maintenance report, and the fit with the operation and maintenance knowledge base. A five-point quantitative scoring system is used to generate specific scoring opinions. During scoring, the review input includes the sample to be verified and several (usually n=2) most similar sub-scenario graphs retrieved from the initial graph. When the total score of the sample is lower than a set threshold, an automated correction process for low-scoring samples is triggered. The large language model combines sample data, scoring opinions, and context to perform targeted revisions and output the corrected sample. During correction, prompts guide the large language model to initiate its automated correction process.

[0081] After startup, the system first constructs a low-scoring sample to be corrected, clarifies the scoring opinions on the shortcomings, ensures that the revision is relevant to the scenario and contextual prompts, and focuses on the requirements of the problems pointed out in the scoring. The prompts are then fed into the large language model, which performs logical deduction and content optimization based on the information in the prompts. Finally, the large language model directly outputs the revised sample, completing the automated correction. Since the aforementioned similarity-based initial screening has already eliminated weakly related samples, the samples entering the correction stage all have sufficient knowledge references. Therefore, automatic correction has high feasibility and success rate in practice.

[0082] Example 2:

[0083] like Figure 2-4 As shown, based on real business Q&A data from operators, 37 high-quality network operation and maintenance fault seed Q&A pairs labeled by experts were selected as input to construct an initial knowledge graph containing 446 entities, 6 types of relations and 526 triples.

[0084] Based on the initial knowledge graph and structured thinking chain prompt template, the three-step process of "analysis-planning-synthesis" is executed using the GPT5 high-performance large language model to generate a total of 101 synthesized data, covering multiple types of operation and maintenance faults, handling processes and tool call scenarios.

[0085] To avoid interference from redundant and irrelevant samples, the Qwen embedding 8b text encoding model was used to calculate the semantic similarity between the synthetic data and the seed samples. Furthermore, a similarity distribution curve and Spearman correlation coefficient statistical features were constructed based on the sub-scene fields. The results are as follows: Figure 2 The results show that among the 97 deduplicated synthetic data, the average Spearman correlation coefficient of 0.8404 was taken as the threshold for judgment, i.e., the red line in the figure. 41 data points below the red line were judged as low similarity samples and removed. The similarity screening results were consistent with the manual judgment in 95%, which verifies the effectiveness of the high-quality data selection mechanism based on similarity distribution proposed in this invention in filtering redundant samples.

[0086] Based on this, logical consistency verification and self-correction experiments were further conducted on the 56 candidate samples after screening.

[0087] The experiment used the GPT5 large language model combined with a custom verification template to score and interpret the samples on a five-point scale, based on four dimensions: the rigor of problem handling logic, the coverage of fault analysis, the compliance of operation and maintenance reports, and the fit with the operation and maintenance knowledge base. The initial scoring results are shown in Table 1, with an average score of 2.78. Among them, 15 samples scored below three points and were automatically identified as objects requiring correction.

[0088] The results of automatic correction and re-scoring are shown in Table 2. All low-scoring samples were improved to three points or above, the average score was raised to 3.20, and the correction success rate reached 100%. The experiment shows that the self-verification and automatic repair mechanism can effectively identify and improve reasoning biases and knowledge defects in synthetic data, thereby significantly improving the logical coherence and knowledge accuracy of synthetic data.

[0089] The corrected, high-quality synthetic data was transformed into triples and added to the initial network operation and maintenance knowledge graph, achieving incremental updates to the graph scale. The number of newly added triples accounted for approximately 131%. Experiments show that the integrated framework of "thinking chain-driven self-planning synthesis + two-stage verification and correction" can achieve high-precision knowledge generation and dynamic updates under low-resource conditions. It not only effectively improves the automation and controllability of knowledge graph construction, but also demonstrates good versatility and engineering feasibility in complex network operation and maintenance scenarios.

[0090] The above are merely preferred embodiments of the present invention and are illustrative in nature, not restrictive. Those skilled in the art will understand that many changes, modifications, and even equivalents can be made within the spirit and scope defined by the claims of the present invention, all of which will fall within the protection scope of the present invention.

Claims

1. A self-programming and self-verification network operation and maintenance knowledge graph generation method based on a thought chain, characterized in that, The network operation knowledge graph generation method specifically comprises the following steps: S1, collate the multi-scenario business data, define the ontology information of the knowledge graph, and construct an initial knowledge graph; S2, first analyze the structure of high-quality seed question and answer pairs and tool tables, then generate a synthesis plan after analysis, and generate synthesis data according to the synthesis plan driving the large prophecy model; S3, select data on the form and distribution levels of the synthesis data, and after self-checking and correcting the synthesis data on the content and structure levels, high-quality synthesis data is obtained; S4, parse the ontology information of the knowledge graph in the high-quality synthesis data string through regular expressions, the ontology information includes triples of typical questions, scenes, sub-scenes, thinking chain steps and tools, and the triples are added to the initial knowledge graph according to the mapping rules to obtain an updated knowledge graph.

2. The thought chain-based self-programming and self-verification network operation and maintenance knowledge graph generation method according to claim 1, characterized in that, The step S1 specifically comprises the following steps: S11, the multi-scenario business data information is obtained by the business experts in the field of network operation technology, and the ontology information of the knowledge graph is defined jointly by the algorithm experts; The multi-scenario business data specifically includes business flowcharts, process tables, typical operation questions and standardized answer templates; S12, the seed question and answer pair data annotated by the business experts and the ontology information used to construct the knowledge graph are sorted out; S13, after the sorting is completed, each seed question and answer pair is parsed through regular expressions, the specific information of the ontology of the knowledge graph in the string is extracted, and then the initial entity-relation triples are generated according to the extracted specific data, and the initial knowledge graph is constructed.

3. The thought chain-based self-programming and self-verification network operation and maintenance knowledge graph generation method of claim 1, wherein, The step S2 specifically comprises the following steps: S21, structure analysis: input several high-quality seed question and answer pairs annotated by business experts and tool tables into a large language model, guide the large language model to extract question patterns, element co-occurrence relationships, typical reasoning paths and tool calling paradigms in the high-quality seed question and answer pairs and tool tables through the templating Prompt of the large language model, and construct a reusable structural element library therefrom; The structural element library includes template definition, parameter dependency mapping and commonly used tool calling templates; S22, synthesis planning: construct the Prompt of the large language model based on the seed data sample, the structural element library and the prompt word template, provide seed data, and drive the large language model to construct the prompt word according to the seed data and the synthesis data to generate an executable synthesis plan; The specific requirements of the synthesis plan for the Prompt are: new sub-scene description, semantic variant strategy, element replacement rule, slot definition and replacement list, and upper limit of the number of generated samples of each type, to control the generation strategy and prevent pattern collapse; S23, synthesis execution: the Prompt of the large language model drives the large language model to batch generate synthesis data according to the seed data sample and the synthesis plan, and the synthesis data output includes structured knowledge graph triple information of typical questions, scenes, sub-scenes, thinking chain steps and tools; When executed, a plurality of groups of simple random sampling with replacement are performed on the high-quality seed question and answer data set to generate seed data samples, and a quantity upper limit is set for each high-quality seed question and answer according to the required generated data quantity, thereby improving sample diversity and reducing homogenization risk.

4. The thought chain-based self-programming and self-verification network operation and maintenance knowledge graph generation method of claim 1, wherein, In the step S3, when high-quality data selection is performed in terms of form and distribution, the high-quality data selector is used to filter out synthetic data close to the semantic feature distribution of the seed data, i.e., to select high-quality synthetic data; During running, each piece of synthetic data text is parsed into a string of "sub-scene-tools" using regular matching, the string is spliced into a piece of comparison text for comparison, and the text is calculated by using a text encoding model to obtain a text embedding vector, that is, the comparison text is mapped to a vector space, and during calculation, the similarity sequence of each piece of synthetic data with all seed samples , Specifically: wherein, representing synthetic data similarity between the nth seed sample; The similarity of all synthetic data to the seed sample is collected to form a full set sample similarity sequence, and the full set sample similarity sequence distribution is fitted to obtain the overall distribution characteristics of the full set sample similarity sequence; Quantifying the consistency of the single sample similarity sequence with the distribution of the whole set sample similarity sequence, using the Spearman correlation coefficient As an indicator of consistency, the retention threshold derived from the population statistic, specifically: wherein, with denote the mean and standard deviation, respectively, of all samples of the respective group, denotes an adjustable empirical coefficient; Samples satisfying the inequality are retained as pre-screening samples, and synthetic data that significantly deviates from the seed sample distribution, is irrelevant to semantics, or is repeated is removed at the statistical level.

5. The thought chain-based self-programming and self-verification network operation and maintenance knowledge graph generation method of claim 1, wherein, In the step S3, when the synthetic data is self-checked and corrected in terms of content and structure, the synthetic data self-checker is used to sequentially perform regularization repair, content quality evaluation, and low-score automatic correction processing on the synthetic data; When the synthetic data self-checker is running, first, the structural consistency of all synthetic data is checked and regularization repair is performed: when regularization repair is performed, first, samples that are inconsistent with the format of the seed data or have inconsistent parameters before and after are located; After positioning, the samples that are inconsistent with the format of the seed data and have inconsistent parameters before and after are automatically processed by calling the rule repair script, realizing structural repair without the participation of a large language model; After the format and parameter problems are repaired, a business expert leads the design of review rules and standards that fit the operation and maintenance scenarios, calls a high-performance large language model to automatically perform content quality evaluation according to a templated Prompt, and quantitatively scores from four operation and maintenance exclusive dimensions, i.e., treatment logic rigor, fault analysis coverage, operation and maintenance report compliance, and operation and maintenance knowledge base fit degree, in a five-point system and generates special scoring opinions; when scoring, the review input contains the sample to be verified, and a number of most similar sub-scenario graphs retrieved from the initial graph; when the sample total score is lower than the set threshold, the low-score sample automatic correction process is triggered, the large language model combines the sample data, scoring opinions, and context to perform directional revision and output the revised sample.

Images