Fault text intelligent processing and classification method and system

Abstract

The invention aims to provide an intelligent fault text processing and classifying method, which realizes automatic identification and statistical analysis of fault types and fault parts through a whole process of text preprocessing, keyword extraction and rule and machine learning mixed classification. In order to improve the effect, a'rule + zero sample learning 'mixed strategy is upgraded into a more complex integrated model with a feedback or self-learning mechanism; a more advanced feature engineering method is introduced, so that the model not only depends on vocabularies.

Description

Technical Field

[0001] This application relates to the field of fault diagnosis and processing, and in particular to a method and system for intelligent processing and classification of fault text. Background Technology

[0002] In equipment operation and maintenance and after-sales repair scenarios, fault logs are typically recorded in text format and stored in structured data files such as Excel. These fault texts often suffer from inconsistent formatting, vague descriptions, and noise (such as URLs, special symbols, or non-string data). Existing fault text processing methods usually rely on manual experience-based classification or simple rule matching based on keywords, which is inefficient and struggles to handle vague descriptions without clear keywords (such as "insufficient equipment power"). Traditional machine learning methods are limited by the difficulty and cost of obtaining labeled data. Currently, there is a lack of efficient, generalized, and accurate intelligent recognition and classification systems specifically designed for fault texts of specific equipment such as "mainframes." Summary of the Invention

[0003] The purpose of this invention is to provide a method for intelligent processing and classification of fault text. This method achieves automatic identification and statistical analysis of fault types and fault locations through a complete process of "text preprocessing → keyword extraction → hybrid classification of rules and machine learning".

[0004] Specific technical solution: A method for intelligent processing and classification of fault text includes the following steps: a. Text preprocessing: Perform standardization cleaning and word segmentation on the fault description text, including converting non-string data into strings and handling null values, removing URLs and special characters, converting to lowercase, using word segmentation tools to segment Chinese words and filter stop words and single characters; b. Keyword extraction: Segment the fault description text and extract words with a length greater than 1 as a keyword list; c. Fault Classification: The text processed in step a is analyzed to identify the fault type and location using a hybrid classification strategy. This hybrid classification includes: i. Rule matching: Based on the preset fault type keyword library and fault location keyword library, the text is precisely matched. If the match is successful, the rule classification result is output. ii. Machine learning classification: A zero-shot classifier is built based on the loaded BERT model to score the semantic similarity between the fault description text and the candidate labels. The candidate labels include a list of fault types and a list of fault locations. iii. Result fusion: The rule classification result of step ci is used first. When the rule classification result is "unknown", the machine learning classification result of step c.ii is used as the final classification result. d. Output of results: Save the fault classification results to a spreadsheet file and generate a visualization of the classification distribution.

[0005] Furthermore, in the text preprocessing, the specific steps for processing non-string input are: converting non-string data into strings, and converting null values ​​NaN into empty strings.

[0006] Furthermore, in the text preprocessing, the specific steps for removing URLs and special characters are: using regular expressions to remove URLs and special characters.

[0007] Furthermore, in the rule matching, the fault type keyword library defines a set of keywords corresponding to the fault type; the fault location keyword library defines a set of keywords corresponding to the fault location.

[0008] Furthermore, in the machine learning classification, the zero-shot classifier selects the label with the highest semantic similarity score that is higher than a preset threshold as the classification result.

[0009] A fault text intelligent analysis system for performing the above method includes: a. Model preparation module: Used to load BERT models and build zero-shot classification pipelines; b. Text preprocessing module: used to perform the standardized cleaning and word segmentation functions of step a in claim 1; c. Keyword extraction module: used to perform the keyword extraction function in step b of claim 1; d. Fault classification module: used to perform the hybrid fault classification function in step c of claim 1; e. Result output module: used to perform the result saving and visualization functions of step d of claim 1.

[0010] To improve performance, the "rule + zero-shot learning" hybrid strategy is upgraded to a more complex ensemble model with feedback or self-learning mechanisms; more advanced feature engineering methods are introduced to make the model not only dependent on vocabulary. Specific improved technical solutions are as follows: A knowledge graph-enhanced intelligent processing and classification method for fault text includes the following steps: a. Model and knowledge preparation: i. Load the pre-trained language model and build a zero-shot classification pipeline; ii. Construct and access a fault domain knowledge graph (KG), which includes fault locations, fault types, and semantic relationships between them; b. Text preprocessing and word embedding: i. Perform standardization cleaning on the original fault description text, including processing non-string data, removing noise, and converting to lowercase; ii. Use the word segmentation technique built into the language model to encode and embed words at the word level to handle out-of-vocabulary words; c. Entity relation extraction and rule matching: i. Utilize the sequence labeling capability of the language model to extract entity relationship triples between fault location entities and fault type entities from the preprocessed text; ii. Based on the preset fault keyword library and the entity relation triples, perform precise rule matching and classification on the text, and output the rule classification results and their confidence C. rule ; d. Knowledge Graph Augmented Zero-Shot Classification KG-ZSL: i. Input the fault description text into the language model, and the model incorporates the semantic association information of the knowledge graph (KG) when generating semantic vectors to enhance domain semantic understanding; ii. Calculate the similarity between the enhanced semantic vector and the candidate labels using the zero-shot classifier, and output the KG-ZSL classification result and its confidence score; the candidate labels include a list of fault types and a list of fault locations. KG-ZSL ; e. Dynamic weight fusion and classification: i. Calculate the dynamic fusion weight α based on the state and confidence level of the rule matching results; ii. Combine the rule-based classification results and the KG-ZSL classification results to generate the final classification result: If the rule matches successfully and the confidence level is C rule Higher than the preset threshold T H If α tends to follow the rule outcome; If the rule matching fails or the confidence level is C rule Below the threshold T L If α tends to be the result of KG-ZSL; If C KG-ZSL If the sample falls within the preset high uncertainty range of 0.45-0.55, then the sample is marked as a high uncertainty sample. f. Active learning feedback: i. Automatically recommend the high-uncertainty samples marked in step e.ii to experts for minor manual annotation; ii. Use the small number of manually labeled samples to fine-tune the language model, update the model parameters, and achieve continuous performance optimization of the system.

[0011] Furthermore, the fault domain knowledge graph (KG) is used in the zero-shot classification KG-ZSL process to enhance the language model's understanding of the semantic relationships between domain concepts.

[0012] Furthermore, in step e, the calculation of the dynamic fusion weight α is based on the success state of rule matching and C. rule and C KG-ZSL The relative size.

[0013] A fault text intelligent analysis system for performing the improved method described above, comprising: a. Knowledge Graph Module: Used to store, manage, and provide the fault domain knowledge graph KG; b. Sequence labeling module: used to perform the entity relationship extraction, i.e., the function of step ci; c. Classification Fusion Module: This module performs the zero-shot classification of the knowledge graph enhancement, i.e., step d, and the dynamic weight fusion, i.e., steps ei-e.ii. d. Active learning module: used to identify the high-uncertainty samples and manage the fine-tuning, i.e., the function of step f.

[0014] The system further includes a fault trend prediction module, which is used to integrate the final classification results into time series analysis to achieve short-term trend early warning for specific fault types or locations.

[0015] Beneficial effects: The intelligent processing and classification method for fault text provided by this invention has the following significant advantages compared with the prior art: It balances accuracy and generalization ability: It adopts a hybrid strategy of "rule priority and machine learning supplementation" to ensure the accuracy of classification, while using zero-shot learning to solve the classification problem of text without clear keywords, thereby improving coverage and generalization ability.

[0016] No labeled training data required: Zero-shot classification is achieved by leveraging BERT's semantic understanding capabilities, avoiding the reliance on large amounts of manually labeled training data in traditional supervised learning and reducing implementation costs.

[0017] High robustness: The text preprocessing module has the ability to unify data types and remove URLs and special characters, enabling the system to process raw Excel fault log data with various noise.

[0018] Interpretability and speed: The rule matching part inherits the advantages of traditional methods, is suitable for texts that explicitly contain keywords, and is fast and the results are easy to interpret.

[0019] Furthermore, this invention also achieves automated, high-precision, and highly generalized classification of structured fault text logs through a composite strategy of "rules + KG-ZSL + active learning," which is particularly suitable for industrial application scenarios lacking labeled data, providing strong technical support for equipment fault analysis, after-sales maintenance record classification, and other tasks. Attached Figure Description

[0020] Figure 1 A flowchart of the intelligent processing and classification method for fault text provided in Embodiment 1 of this application; Figure 2 The flowchart shows the intelligent processing and classification method for fault text based on knowledge graph enhancement provided in Embodiment 2 of this application. Detailed Implementation

[0021] Exemplary embodiments will be described in detail herein. The implementations described in the following exemplary embodiments do not represent all implementations consistent with this application. The terminology used in this application is for the purpose of describing particular embodiments only and is not intended to be limiting of this application. The singular forms “a,” “the,” and “the” used in this application are also intended to include the plural forms unless the context clearly indicates otherwise.

[0022] The following specific embodiments are given to illustrate the technical solution of this application in detail.

[0023] Example 1: A method for intelligent processing and classification of fault text is mainly used to automatically process and classify fault description text in Excel spreadsheets (identifying fault type and fault location) and generate visualization results.

[0024] The core objective is to automatically identify the type of fault (e.g., "damage" or "leakage") and the location of the fault (e.g., "engine" or "working device") from structured Excel data (including fault descriptions) using techniques such as text cleaning, keyword matching, and machine learning (zero-shot classification). The final output is the classification results and a visualization of the distribution. It is suitable for scenarios such as equipment fault analysis and after-sales maintenance record classification.

[0025] Define stop words (for text cleaning), fault type keyword library (such as "fault", "damage", etc.), and fault location keyword library (such as "engine", "circuit", etc.).

[0026] Text preprocessing: Specify the column name in Excel that stores the fault description (e.g., "Fault Description") to ensure that subsequent processing can locate the target data.

[0027] The core methods of text cleaning include: converting non-string data into strings (handling empty values ​​or numbers); removing URLs and special characters, and converting to lowercase; using jieba for word segmentation, filtering stop words and single characters to obtain the cleaned word segmentation results; and extracting keywords (words with a length > 1) from the text for subsequent analysis.

[0028] Fault classification logic: The tool offers three classification methods, the core of which is to identify the "fault type" and "fault location":

[0029] The final results are saved as an Excel file (fault classification results.xlsx) for easy subsequent analysis.

[0030] There is no existing technology specifically for mainframe fault text recognition. A fault text intelligent analysis system aims to automatically identify and statistically analyze fault types and locations from fault logs in Excel format through a complete process of "text preprocessing → keyword extraction → fault classification (rules + machine learning) → result visualization".

[0031] I. Model Preparation Module: Local Caching and Automatic Download Mechanism 1. Based on the loaded BERT model, construct a "zero-shot classification" pipeline to provide a foundation for subsequent machine learning classification.

[0032] II. Text Preprocessing Module: Standardized Cleaning and Word Segmentation Core Functions The original fault description (which may contain noise and be in non-string format) is converted into structured, analyzable text data, paving the way for subsequent feature extraction and classification.

[0033] Key Algorithms and Steps 1. Unified data types: Handling non-string input (such as numbers, NaN): Convert non-string input to strings, and convert NaN to an empty string to avoid errors in subsequent functions.

[0034] 2. Noise Removal: Use regular expressions to remove URLs (`http\S+`) and special characters (`[^\w\s]`), keeping only text and spaces; Convert the text to lowercase to eliminate case differences (e.g., unify "Leak" and "leak" into lowercase).

[0035] 3. Chinese word segmentation and stop word filtering: Use `jieba` (a Chinese word segmentation tool) to segment the cleaned text (e.g., "hydraulic valve leakage" is segmented into "hydraulic valve" and "leakage"). Filter out stop words (such as "de", "le", "zai" in Chinese) and single characters (words with a length ≤ 1), and retain meaningful words to reduce redundant features.

[0036] III. Keyword Extraction Module: Screening of Effective Information Based on Word Segmentation Core Function Extract keyword vocabulary from the original fault description to provide intuitive information support for manual analysis or subsequent feature engineering.

[0037] Algorithm Logic Directly use `jieba` to segment the original text (without lowercasing, retaining the original semantics), only filter single characters (only words with a length > 1 are retained), and finally output a list of keywords for each fault description (e.g., "engine shakes severely" → ["engine", "shake", "severely"]).

[0038] IV. Fault Classification Module: Hybrid Scheme of Rules + Machine Learning This is the core module of the system. Through the strategy of "rules first, machine learning for supplementation", it realizes the automatic classification of fault types (such as leakage, blockage) and fault locations (such as engine, oil circuit), taking into account both accuracy and generalization ability.

[0039] 4.1 Classification Based on Keyword Rules (Rule - Based) Core Logic Use the preset "fault keyword library" for exact matching, which is suitable for processing fault descriptions that clearly contain keywords, with high speed and strong interpretability.

[0040] Algorithm Steps 1. Definition of Keyword Library: Pre - define two types of keyword dictionaries: `FAULT_TYPE_KEYWORDS`: Fault types (e.g., for "leakage", the corresponding keywords are ["leakage", "seepage", "oil leak"]); `FAULT_LOCATION_KEYWORDS`: Fault locations (e.g., for "engine", the corresponding keywords are ["engine", "engine", "motor"]).

[0041] 2. Text Matching: Traverse the fault description text. If it contains any keyword of a certain category, then take that category as the classification result (e.g., if the text contains "oil leak" → match the "leakage" type); If no keyword is matched, the classification result is set to "unknown".

[0042] 4.2 Machine Learning-Based Classification (ML-Based: Zero-Shot Learning) Core Logic For vague fault descriptions without keywords (such as "insufficient equipment power"), the semantic understanding capability of the BERT pre-trained model is utilized to achieve classification through "zero-shot learning," and its generalization ability is superior to rule-based methods.

[0043] Algorithm Principles 1. Zero-Shot Classification: Without needing to label training data, the "fault type / location list" can be directly used as candidate labels, allowing the model to score based on the similarity between the text semantics and the labels; For example, compare "insufficient equipment power" with labels such as "performance degradation" and "blockage", and select the label with the highest similarity (score > threshold 0.5) as the result.

[0044] 2. The role of the BERT model: The `bert-base-chinese` model converts Chinese text into semantic vectors, capturing the semantic association between "insufficient power" and "performance degradation," and achieving cross-keyword classification.

[0045] 3. Threshold filtering: Set a score threshold (0.5 for both type and location). If the highest score is below the threshold, set the classification result to "unknown" to avoid low-confidence prediction.

[0046] 4.3 Hybrid Classification Core Strategy Rule priority: If the rule classification has yielded a clear result (not "unknown"), use the rule result directly (higher accuracy). Machine learning fill-in: If the rule classification result is "unknown" (e.g., the text has no keywords), then the result of machine learning (zero samples) is used as the final classification to cover scenarios not covered by the rule.

[0047] Example 2: To improve performance, the "rule + zero-shot learning" hybrid strategy is upgraded to a more complex ensemble model with feedback or self-learning mechanisms; more advanced feature engineering methods are introduced to make the model not only dependent on vocabulary. Specific improved technical solutions are as follows: This specific implementation describes in detail how to implement an intelligent processing and classification method for fault text based on knowledge graph enhancement and active learning feedback.

[0048] 1. System Architecture and Module Division This system adopts a modular design and mainly includes: knowledge preparation and model loading module, text preprocessing and embedding module, entity relationship extraction module, hybrid classification module (including rule matching and KG-ZSL), dynamic fusion module, active learning feedback module, and result output module.

[0049] 2. Knowledge Preparation and Model Loading 2.1 Language Model Loading The system loads a pre-trained Chinese language model, such as BERT-base-chinese. A zero-shot classification pipeline is then built based on this model, serving as the foundation for machine learning classification.

[0050] 2.2 Construction of Fault Domain Knowledge Graph (KG) Construct a fault domain knowledge graph (KG) to store fault entities (such as location, type, and part) and their interrelationships.

[0051] Entities include the location of the fault (e.g., "engine", "fuel line") and the type of fault (e.g., "leakage", "damage").

[0052] Relations: Defines the professional relationships between entities (e.g., "engine" is prone to "vibration"; "oil pipe" belongs to "oil circuit").

[0053] The KG will be used to enhance the BERT model's understanding of specialized fault semantics, especially in the zero-shot classification stage.

[0054] 3. Text preprocessing and word embedding 3.1 Standardized Cleaning The steps involved in cleaning the original text include: Data type unification: Convert non-string data (such as numbers or empty values ​​NaN) in Excel to strings, and convert NaN to an empty string.

[0055] Noise removal: Use regular expressions to remove URLs (such as http\S+) and special characters (such as [^\w\s]) from the text, keeping only the text and spaces.

[0056] Lowercase conversion: Converts all text to lowercase, eliminating the case difference.

[0057] 3.2 Subword Embedding Instead of using traditional jieba segmentation (which relies on a dictionary), this implementation uses the sub-word splitting technology built into the BERT model (such as WordPiece or BPE). The cleaned text is split into sub-word tokens, which are then fed into the BERT model to generate sub-word-level embedding vectors. This effectively improves the system's ability to handle out-of-vocabulary words and specialized new words in faulty texts.

[0058] 4. Entity Relationship Extraction and Rule Classification 4.1 Entity-Relation Extraction Use a pre-trained BERT model for sequence labeling to identify key entities in the text (such as location and type) and explicitly extract fault entity relationship triples (e.g., extract (engine, fault type, vibration) from "severe engine vibration").

[0059] This triple serves as input for enhanced rule matching, improving the accuracy and binding force of the rules.

[0060] 4.2 Rule-Based Matching Matching is performed based on a predefined keyword library and extracted entity relation triples: Keyword database: Define FAULT_TYPE_KEYWORDS and FAULT_LOCATION_KEYWORDS.

[0061] Matching logic: The rule is considered to be successfully matched only when both part entity keywords and type entity keywords (or their synonyms) are matched in the text, and the relation triple is reasonable. The rule classification result Rrule is then output.

[0062] Rule confidence level Crule: If an exact match is found, Crule is set to a high confidence value (e.g., 0.95); if the match fails, the result Rrule is set to "unknown" and Crule is set to 0.

[0063] 5. Knowledge Graph Augmented Zero-Shot Classification (KG-ZSL) For texts where rule matching fails (i.e., Rrule is "unknown"), KG-ZSL is activated to perform supplementary classification.

[0064] Knowledge Graph Enhancement: When the BERT model generates text semantic vectors, it interacts with a pre-loaded fault domain knowledge graph (KG) to integrate professional association information between fault entities into the text embedding. This enables the model to better understand semantic associations such as the type of "oil leak" and "leakage," and the location of "oil line."

[0065] Similarity scoring: Using enhanced semantic vectors, calculate the semantic similarity score between the text and candidate tags (such as "leak", "blockage", "engine").

[0066] Output: Select the label with the highest score as the R result. KG-ZSL The similarity score is the confidence level C. KG-ZSL .

[0067] 6. Dynamic weight fusion and active learning feedback 6.1 Dynamic Weight Fusion The system no longer simply executes "failure and switch," but instead makes dynamic decisions based on confidence level C: Rule priority: If Crule > 0.9, then the final classification result R is given. final = Rrule.

[0068] Intelligent fill: If Crule = 0 (rule matching failed), then the final result R... final = R KG-ZSL However, it must satisfy C. KG-ZSL Exceeding the minimum threshold Tmin (e.g., 0.5).

[0069] High uncertainty judgment: If C KG-ZSL Located in the preset high uncertainty interval T AL (For example, 0.45 ≤ C) KG-ZSL If the value is ≤0.55, then the sample is marked as a "high uncertainty sample" and R is set. final For R KG-ZSL (or "Pending manual review").

[0070] 6.2 Active Learning Feedback This is the key mechanism for achieving continuous system optimization.

[0071] Sample Recommendation: The system automatically pushes all the "high uncertainty samples" marked in step 6.1 to the expert review interface. These samples are the most "confusing" for the model and are the most valuable for improving model performance.

[0072] Manual annotation and fine-tuning: After experts accurately annotate a small number of high-uncertainty samples, they use these high-value small batches of data to perform incremental fine-tuning on the BERT model.

[0073] Iterative loop: The fine-tuned model parameters will update the zero-shot classification pipeline, thereby improving the accuracy and robustness of the system when processing similar fuzzy text, and achieving continuous improvement in system performance.

[0074] 7. Results Output and Application Final classification result R final (Including fault type and location) The results are saved to an Excel file, and a visual chart of fault distribution is generated. Furthermore, the system can integrate the results with a fault trend prediction module, using time series analysis to provide early warnings of increasing fault frequencies in specific locations or types.

[0075] The above description is merely a preferred embodiment of this application and is not intended to limit this application. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the scope of protection of this application.

Claims

1. A method for intelligent processing and classification of fault text, characterized in that, Includes the following steps: a. Text preprocessing: Perform standardization cleaning and word segmentation on the fault description text, including converting non-string data into strings and handling null values, removing URLs and special characters, converting to lowercase, using word segmentation tools to segment Chinese words and filter stop words and single characters; b. Keyword extraction: Segment the fault description text and extract words with a length greater than 1 as a keyword list; c. Fault Classification: The text processed in step a is analyzed to identify the fault type and location using a hybrid classification strategy. This hybrid classification includes: i. Rule matching: Based on the preset fault type keyword library and fault location keyword library, the text is precisely matched. If the match is successful, the rule classification result is output. ii. Machine learning classification: A zero-shot classifier is built based on the loaded BERT model to score the semantic similarity between the fault description text and the candidate labels. The candidate labels include a list of fault types and a list of fault locations. iii. Result fusion: The rule classification result of step ci is used first. When the rule classification result is "unknown", the machine learning classification result of step c.ii is used as the final classification result. d. Output of results: Save the fault classification results to a spreadsheet file and generate a visualization of the classification distribution.

2. The method as described in claim 1, characterized in that, In the text preprocessing, the specific steps for processing non-string input are: converting non-string data into strings and converting null values ​​NaN into empty strings.

3. The method as described in claim 1, characterized in that, In the text preprocessing, the specific steps for removing URLs and special characters are: using regular expressions to remove URLs and special characters.

4. The method as described in claim 1, characterized in that, In the rule matching, the fault type keyword library defines a set of keywords corresponding to the fault type; the fault location keyword library defines a set of keywords corresponding to the fault location.

5. The method as described in claim 1, characterized in that, In the machine learning classification, the zero-shot classifier selects the label with the highest semantic similarity score that is higher than a preset threshold as the classification result.

6. A fault text intelligent analysis system for performing the method according to any one of claims 1 to 5, characterized in that, include: a. Model preparation module: Used to load BERT models and build zero-shot classification pipelines; b. Text preprocessing module: used to perform the standardized cleaning and word segmentation functions of step a in claim 1; c. Keyword extraction module: used to perform the keyword extraction function in step b of claim 1; d. Fault classification module: used to perform the hybrid fault classification function in step c of claim 1; e. Result output module: used to perform the result saving and visualization functions of step d of claim 1.

7. A method for intelligent processing and classification of fault text based on knowledge graph enhancement, characterized in that, Includes the following steps: a. Model and knowledge preparation: i. Load the pre-trained language model and build a zero-shot classification pipeline; ii. Construct and access a fault domain knowledge graph (KG), which includes fault locations, fault types, and semantic relationships between them; b. Text preprocessing and word embedding: i. Perform standardization cleaning on the original fault description text, including processing non-string data, removing noise, and converting to lowercase; ii. Use the word segmentation technique built into the language model to encode and embed words at the word level to handle out-of-vocabulary words; c. Entity relation extraction and rule matching: i. Utilize the sequence labeling capability of the language model to extract entity relationship triples between fault location entities and fault type entities from the preprocessed text; ii. Based on the preset fault keyword library and the entity relation triples, perform precise rule matching and classification on the text, and output the rule classification results and their confidence C. rule ; d. Knowledge Graph Augmented Zero-Shot Classification KG-ZSL: i. Input the fault description text into the language model, and the model incorporates the semantic association information of the knowledge graph (KG) when generating semantic vectors to enhance domain semantic understanding; ii. Calculate the similarity between the enhanced semantic vector and the candidate label using the zero-shot classifier, and output the KG-ZSL classification result and its confidence score; Candidate labels include a list of fault types and a list of fault locations. KG-ZSL ; e. Dynamic weight fusion and classification: i. Calculate the dynamic fusion weight α based on the state and confidence level of the rule matching results; ii. Combine the rule-based classification results and the KG-ZSL classification results to generate the final classification result: If the rule matches successfully and the confidence level is C rule Higher than the preset threshold T H If α tends to follow the rule outcome; If the rule matching fails or the confidence level is C rule Below the threshold T L If α tends to be the result of KG-ZSL; If C KG-ZSL If the sample falls within the preset high uncertainty range of 0.45-0.55, then the sample is marked as a high uncertainty sample. f. Active learning feedback: i. Automatically recommend the high-uncertainty samples marked in step e.ii to experts for minor manual annotation; ii. Use the small number of manually labeled samples to fine-tune the language model, update the model parameters, and achieve continuous performance optimization of the system.

8. The method as described in claim 7, characterized in that, The fault domain knowledge graph (KG) is used to enhance the language model's understanding of semantic relationships between domain concepts during zero-shot classification (KG-ZSL).

9. The method as described in claim 7, characterized in that, In step e, the calculation of the dynamic fusion weight α is based on the success status of rule matching and C. rule and C KG-ZSL The relative size.

10. A fault text intelligent analysis system, used to perform the method according to any one of claims 7 to 9, characterized in that, include: a. Knowledge Graph Module: Used to store, manage, and provide the fault domain knowledge graph KG; b. Sequence labeling module: used to perform the entity relationship extraction, i.e., the function of step ci; c. Classification Fusion Module: This module performs the zero-shot classification of the knowledge graph enhancement, i.e., step d, and the dynamic weight fusion, i.e., steps ei-e.ii. d. Active learning module: used to identify the high-uncertainty samples and manage the fine-tuning, i.e., the function of step f.

Images