Converter transformer fault case structured processing method, system, device and medium

Abstract

The invention relates to the technical field of case data structuring, and discloses a structured processing method, system, equipment and medium for a converter transformer fault case, and the method comprises the steps: carrying out the data cleaning and format unification of original text data of the converter transformer fault case, and obtaining text data; preprocessing the text data, and inputting a feature extraction model to obtain a text feature vector, the feature extraction model being constructed based on a natural language processing model; and classifying the text feature vectors through a clustering analysis algorithm, and generating a structured text corresponding to the converter transformer fault case according to a classification result and a fixed label format. According to the method, feature extraction is performed on the converter transformer fault case through a natural language processing technology, so that the integrity rate of feature extraction is improved; through the label distribution method based on clustering analysis, the accuracy of label distribution is improved, high-quality structured text conversion is realized, and accurate and efficient data support is provided for intelligent analysis of converter transformer fault cases.

Description

Technical Field

[0001] The present invention relates to the technical field of case data structuring, and in particular to a method, system, device and medium for structured processing of converter transformer fault cases. Background Art

[0002] As the core equipment of UHVDC transmission systems, converter transformers perform the dual functions of AC / DC power conversion and voltage level regulation. Their structural complexity is reflected in the highly coupled nature of multiple subsystems, including bushing connections, winding arrangement, insulation design, and cooling systems. Furthermore, differences in design specifications and process standards between domestic and international manufacturers lead to significant engineering feature dependency and manufacturer-specific failure modes.

[0003] The current power operation and maintenance sector has accumulated a massive amount of text data on converter transformer fault cases, covering the entire life cycle of equipment, including factory testing, operational anomalies, fault diagnosis, and maintenance records. Studying and analyzing converter transformer fault cases is an effective way to improve converter transformer reliability. With the development of machine learning technology, machine learning can now be used to analyze fault cases. Although this type of data provides a potential knowledge base for reliability optimization through machine learning, technical bottlenecks still exist in practical applications: most current fault cases are recorded in the form of free text, which suffers from problems such as redundant expression, information fragmentation, and noise interference, making semantic parsing difficult; traditional text matching methods rely on keyword rule libraries, which are difficult to deal with the polysemy of professional terms and contextual association features, and cannot achieve automatic decoupling of multi-feature coupling scenarios in power systems; in addition, existing analysis technologies focus on a single dimension and lack the ability to mine horizontal knowledge of fault cases across projects and manufacturers. The heterogeneity of unstructured text and structured monitoring data hinders the fusion of multi-source information.

[0004] Although machine learning technology provides a new path for fault pattern recognition and prediction, it is limited by the lack of structured text data. Existing methods find it difficult to build an intelligent analysis framework covering semantic features, operating parameters, and equipment topology. Therefore, there is an urgent need for a natural language processing method to achieve structured conversion of fault case text. Summary of the Invention

[0005] In order to solve the above technical problems, the present invention provides a structured processing method, system, device and medium for converter transformer fault cases, which converts the unstructured text of converter transformer fault cases into clear and orderly structured data, providing reliable data support for the intelligent analysis of converter transformer fault cases.

[0006] In a first aspect, the present invention provides a structured processing method for converter transformer fault cases, the method comprising:

[0007] Clean and format the original text data of converter transformer fault cases to obtain text data;

[0008] Preprocessing the text data and inputting the preprocessed text data into a preset feature extraction model to obtain a text feature vector, wherein the feature extraction model is constructed based on a natural language processing model, and the preprocessing includes term segmentation and length unification;

[0009] The text feature vectors are classified by a cluster analysis algorithm, and structured text corresponding to the converter transformer fault case is generated according to the classification result and a preset fixed label format.

[0010] Furthermore, the step of cleaning and formatting the original text data of the converter transformer fault case to obtain the text data includes:

[0011] Perform data cleaning on the original text data of converter transformer fault cases;

[0012] According to preset modules, the cleaned original text data is divided into modules to obtain text data in a unified format. The preset modules include a problem description module, a cause analysis module, a treatment measure module and a treatment result module.

[0013] Furthermore, the step of preprocessing the text data includes:

[0014] Dynamically segmenting professional terms in the electric power field in the text data using a byte pair encoding algorithm;

[0015] The length of the initial text segments obtained by dynamic word segmentation is unified to obtain text segments.

[0016] Furthermore, the feature extraction model is constructed based on the RoBERTa model, the Adam optimizer is used to tune the model parameters, and the cross entropy loss is used as the loss function.

[0017] Furthermore, the step of classifying the text feature vectors by a cluster analysis algorithm and generating a structured text corresponding to the converter transformer fault case according to the classification result and a preset fixed label format includes:

[0018] Classifying the text feature vectors using a K-means clustering algorithm to obtain a classification result;

[0019] According to the classification result and a preset fixed label format, the text feature vector is filled into the corresponding label field to obtain a structured text of the converter transformer fault case;

[0020] The tag fields include project name, manufacturer, product type, product grade, problem type, problem location, problem cause and specific problem situation.

[0021] Furthermore, the step of classifying the text feature vectors by a cluster analysis algorithm and generating a structured text corresponding to the converter transformer fault case according to the classification result and a preset fixed label format includes:

[0022] Performing dimensionality reduction on the text feature vector, using a noise application space clustering algorithm to identify a core area of ​​the reduced text feature vector, and using a hierarchical density clustering algorithm to hierarchically divide the identified core area to obtain a plurality of subclusters;

[0023] Calculate the label similarity between each subcluster and each label field in the fixed label format, and determine the cluster label corresponding to each subcluster based on the label similarity. The label fields include project name, manufacturer, product type, product grade, problem type, problem location, problem cause, and specific problem situation.

[0024] According to the cluster label, each sub-cluster is filled into the corresponding label field to obtain a structured text corresponding to the converter transformer fault case.

[0025] Furthermore, the step of determining the cluster label corresponding to each sub-cluster according to the label similarity includes:

[0026] Determining, based on the label similarities, whether each subcluster satisfies a conflict resolution condition, wherein the conflict resolution condition includes that at least two label similarities are greater than a first threshold, or that all label similarities are less than a second threshold, and the first threshold is greater than the second threshold;

[0027] In response to the existence of a sub-cluster that meets the conflict resolution condition, the sub-cluster that meets the conflict resolution condition is used as a sub-cluster to be optimized;

[0028] A label co-occurrence graph is constructed using each label field in the fixed label format as a node and the label co-occurrence frequency in the historical fault case library as the edge weight.

[0029] Using a graph embedding algorithm to perform node embedding calculation on the tag co-occurrence graph to obtain a node embedding vector corresponding to each node in the tag co-occurrence graph;

[0030] Calculate the embedding similarity between the center vector of the subcluster to be optimized and the node embedding vector, and determine the cluster label corresponding to the subcluster to be optimized according to the embedding similarity.

[0031] In a second aspect, the present invention provides a structured processing system for converter transformer fault cases, the system comprising:

[0032] A data processing module is used to clean and unify the format of the original text data of the converter transformer fault case to obtain text data;

[0033] A feature extraction module is used to preprocess the text data and input the preprocessed text data into a preset feature extraction model to obtain a text feature vector, wherein the feature extraction model is constructed based on a natural language processing model, and the preprocessing includes term segmentation and length unification;

[0034] The text generation module is used to classify the text feature vectors by using a cluster analysis algorithm, and generate structured text corresponding to the converter transformer fault case according to the classification result and a preset fixed label format.

[0035] In a third aspect, an embodiment of the present invention further provides a computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the steps of the above method when executing the computer program.

[0036] In a fourth aspect, an embodiment of the present invention further provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program implements the steps of the above method when executed by a processor.

[0037] The present invention provides a structured processing method, system, device, and medium for converter transformer fault case analysis. This method uses natural language processing technology to extract features from converter transformer fault cases, resolving the feature extraction challenges caused by redundant text information and diverse expressions of specialized terminology within fault case texts, thereby improving the completeness of feature extraction. Furthermore, a label assignment method based on cluster analysis improves label assignment accuracy and enables high-quality structured text conversion, providing precise and efficient data support for intelligent analysis of converter transformer fault cases. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] Figure 1 1 is a flow chart of a structured processing method for a converter transformer fault case according to an embodiment of the present invention;

[0039] Figure 2 1 is a schematic diagram of a structured processing system for converter transformer fault cases according to an embodiment of the present invention;

[0040] Figure 3 is an internal structural diagram of a computer device according to an embodiment of the present invention;

[0041] Reference numerals:

[0042] 10. Data processing module; 20. Feature extraction module; 30. Text generation module. DETAILED DESCRIPTION

[0043] To make the objectives, technical solutions, and advantages of the embodiments of the present invention more clear, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts shall fall within the scope of protection of the present invention.

[0044] See also Figure 1 A structured processing method for a converter transformer fault case proposed in a first embodiment of the present invention includes steps S10 to S30:

[0045] Step S10, performing data cleaning and formatting on the original text data of the converter transformer fault case to obtain text data;

[0046] Step S20, preprocessing the text data and inputting the preprocessed text data into a preset feature extraction model to obtain a text feature vector, wherein the feature extraction model is constructed based on the RoBERTa model;

[0047] Step S30 , classifying the text feature vectors by a cluster analysis algorithm, and generating a structured text corresponding to the converter transformer fault case according to the classification result and a preset fixed label format.

[0048] The present invention provides a method for extracting and labeling converter transformer fault cases based on natural language processing technology, converting unstructured fault case text into structured text. Due to the inconsistent writing styles and numerous professional terminology aliases in various fault cases in current power systems, the text format of fault cases is not standardized and contains excessive interference information, which greatly complicates case analysis. Therefore, the present invention structures fault cases, enabling users to quickly locate relevant cases based on specific keywords or attributes. This significantly reduces the time cost of information search, improves work efficiency, and enables efficient retrieval and query. Furthermore, the structured data format facilitates the use of data mining algorithms to discover potential patterns, influencing factors, and development trends in fault occurrence, providing a scientific basis for fault warning and risk assessment. Through in-depth analysis of historical fault cases, frequently faulty links or components can be identified, enabling the development of targeted preventive maintenance plans, reducing unplanned downtime, improving equipment reliability and grid stability, and optimizing operation and maintenance strategies. Therefore, structuring fault cases is highly necessary.

[0049] The original text data of converter transformer fault cases in the present invention first needs to be cleaned and formatted. The specific steps include:

[0050] Perform data cleaning on the original text data of converter transformer fault cases;

[0051] According to preset modules, the cleaned original text data is divided into modules to obtain text data in a unified format. The preset modules include a problem description module, a cause analysis module, a treatment measure module and a treatment result module.

[0052] In this embodiment, for converter transformer fault cases written in natural language, data cleaning is first performed to remove irrelevant content from the text, including timestamps and irrelevant identifiers. This also removes irrelevant unstructured information, such as technician comments, extraneous spaces, and line breaks. Electrical professionals then organize the text of each converter transformer fault case into a unified format, specifically including four modules: problem description, cause analysis, treatment measures, and treatment results, to prevent subsequent processing from being affected by inconsistent formatting.

[0053] For the text data obtained after data cleaning and format unification, feature extraction is performed through a preset feature extraction model in this embodiment, wherein the feature extraction model is constructed using a natural language processing model, such as a natural language processing model such as a BERT (Bidirectional Encoder Representations from Transformers) model and a GTP (Generative Pre-trained Transformer) model. Preferably, this embodiment uses a RoBERTa (Robustly Optimized BERT Approach) model to construct a feature extraction model. Compared with other natural language processing models, the RoBERTa model adopts a deeper text understanding mechanism, which can capture rich semantic information in a wider context. The converter transformer fault case text usually contains complex technical terms, professional keywords and long sentence structures. Therefore, using the RoBERTa model to process the converter transformer fault case text has significant advantages in the accuracy of feature extraction and semantic understanding.

[0054] Before inputting text data into the feature extraction model, the text data needs to be preprocessed. The specific steps include:

[0055] Dynamically segmenting professional terms in the electric power field in the text data using a byte pair encoding algorithm;

[0056] The length of the initial text segments obtained by dynamic word segmentation is unified to obtain text segments.

[0057] In this embodiment, the preprocessing step includes text segmentation. The goal of text segmentation is to split long text into small segments suitable for processing by the RoBERTa model, ensuring that each text segment has complete semantic information and meets the model's requirements for input length. The entire text segmentation process can be divided into a text segmentation stage and a filling and truncation stage. Specifically, text segmentation is the process of converting text from an original character sequence into units that the model can process, usually "words" or "subwords". It should be noted here that although the RoBERTa model also has a built-in segmentation algorithm that can split the original text into subword units, due to the presence of a large number of unique and complex terms in the power field, such as "low-end bushing discharge" and "hydrogen cooling system leakage", they may be split into unreasonable subwords, resulting in semantic distortion. Therefore, this embodiment uses segmentation enhancement in the preprocessing stage to improve the domain adaptability of subsequent models for segmentation processing.

[0058] In this embodiment, a Byte Pair Encoding (BPE) algorithm is used for word segmentation during the preprocessing process. During the preprocessing word segmentation, professional terms are protected by custom BPE rules. The custom BPE rules here include a predefined term list and a dynamic expansion word list. The terms included in the predefined term list (such as "low-end bushing discharge") will be forcibly retained to avoid splitting. The dynamic expansion word list includes words that have not been included in the previous word segmentation data table, such as the names of new equipment. The dynamic expansion word list is dynamically added based on the fault case library. Pre-optimization through pre-processing word segmentation can protect the integrity of key terms and achieve fine-grained sub-word splitting when the subsequent input model is input for word segmentation processing.

[0059] After word segmentation, the length of the text may vary. To ensure that the input text has a consistent length in the model, this embodiment uses padding or truncation to achieve length uniformity. Specifically, for shorter texts, special padding markers are used to pad them to a fixed length. For longer texts, the model truncates them, retaining only the first part. This approach avoids problems caused by inconsistent text lengths.

[0060] In addition to preprocessing the text, the dataset used for training the feature extraction model in this embodiment can also be selected from the preprocessed text. Specifically, a certain amount of preprocessed text data is selected and annotated by professionals. The quality of the annotated data directly affects the training effect of the model, so it is necessary to ensure the accuracy and consistency of the annotation. It should be noted that padding during length unification does not affect the training of the model, because the marked parts are not used to calculate the loss function, and the model will ignore the impact of these parts.

[0061] During the training process, this embodiment uses a pre-trained RoBERTa model and fine-tunes it. The purpose of fine-tuning is to adapt the pre-trained model to the specific task of converter transformer failure case text, so further training is performed based on this specific data. Furthermore, the Adam optimizer is used to optimize the model parameters. The Adam optimizer is an adaptive gradient descent method that effectively adjusts the learning rate of each parameter, thereby accelerating convergence. This embodiment uses the cross-entropy loss function to measure the difference between the model output and the true label. By minimizing the loss function, the model continuously optimizes its parameters to better fit the training data. The goal of model training is to improve accuracy by maximizing the probability of the feature class predicted by the model. For example, in the "product type" classification task, the RoBERTa model predicts the product type in each converter transformer failure case based on the context of terms such as "transformer" and "electrical equipment" in the training data. During model training, model performance is monitored using a validation set. The validation set is a portion of the labeled data, and the model is evaluated on the validation set after each round of training to determine the current model performance. In addition to accuracy, this embodiment also uses the F1 score metric to comprehensively evaluate the model's performance on the specific task. The F1 score metric is suitable for cases with class imbalance and can better reflect the model's recognition performance across different categories. For the specific model training steps, refer to the conventional model training steps and will not be detailed here.

[0062] After model training is complete, the feature extraction model can be used to extract features from the preprocessed text data. Based on the extracted text feature vectors, the labels corresponding to the features are determined. However, in a fault case, a single text segment may correspond to descriptions of multiple feature categories. Therefore, clustering is required to further optimize the feature extraction results, thereby generating structured fault case text in a fixed-label format that includes various features.

[0063] In a preferred embodiment, the steps of classifying the text feature vectors using a cluster analysis algorithm and generating structured text corresponding to the converter transformer fault case based on the classification results and a preset fixed label format include:

[0064] Classifying the text feature vectors using a K-means clustering algorithm to obtain a classification result;

[0065] According to the classification result and a preset fixed label format, the classified text feature vector is filled into the corresponding label field to obtain a structured text of the converter transformer fault case;

[0066] The tag fields include project name, manufacturer, product type, product grade, problem type, problem location, problem cause and specific problem situation.

[0067] In this embodiment, since the converter transformer fault case text often involves complex technical descriptions, it contains content from multiple feature categories. For example, a fault case may simultaneously describe multiple features such as project name, manufacturer information, product type, problem type, problem location, and problem cause. These features are often intertwined, and there is a certain degree of ambiguity and complex contextual correlation, making it difficult to directly extract a single category of information from the text. Therefore, this embodiment uses a clustering method to optimize the feature extraction results. Preferably, the K-Means clustering method is used to automatically divide text data containing multiple feature categories into different categories. In this way, clustering can effectively identify the potential structure in the text and group similar texts together, thereby optimizing the effect of feature extraction.

[0068] When applying the K-Means clustering method to converter transformer fault case text, it's first necessary to ensure that the model's feature extraction is complete. The RoBERTa model, through deep semantic understanding, generates a text feature vector for each fault case. These feature vectors are then processed using K-Means clustering to identify different feature categories and optimize the extraction results. The goal of clustering is to automatically assign text to clusters based on semantic similarity. The text within each cluster has high similarity, thus better reflecting the commonalities of features within that category. Initially, a K value (i.e., the number of clusters) of 8 is selected to ensure the stability and accuracy of the clustering results. Next, the K-Means algorithm is used to cluster the feature vectors, outputting the center of each cluster and the distribution of text within the cluster. By calculating the Euclidean distance from each feature vector to the cluster center, features are assigned to the closest cluster center. Based on the assignment results, the average position of each cluster center is recalculated and the cluster center coordinates are updated. This process is repeated until the cluster centers stabilize or meet the specified convergence criteria, completing the feature grouping.

[0069] After clustering is completed, the clustering results need to be analyzed in detail. The text in each cluster represents a feature category, and the cluster center represents the central position of the cluster feature. Through in-depth analysis of the text within the cluster, the representative features of each cluster can be extracted. By analyzing the high-frequency words within the cluster, the feature category represented by each cluster can be determined. For example, a cluster may contain keywords such as "overheating" and "short circuit", then the cluster is likely to represent the "fault type" feature. At the same time, analyze the similarity and consistency of the text within the cluster to ensure the rationality of the clustering results. If the semantic differences of the text within a cluster are large, it may be necessary to adjust the K value or further optimize the feature selection.

[0070] The feature extraction results after clustering optimization will eventually be used to generate standardized labels. In this embodiment, the fixed label format is: project name-manufacturer-product type-product grade-problem type-problem location-problem cause-specific problem situation. According to the fixed label format and the results of clustering analysis, the keywords of the clustered feature vector are filled into the corresponding label field to generate formatted text data. For example, if a text segment belongs to the "project name" cluster, then the "project name" field in its label format will be filled with the corresponding project name; if a text segment belongs to the "problem type" cluster, then the "problem type" field in its label will be filled with the corresponding fault problem type. Through the clustering analysis method, multiple related feature categories can be classified into the same cluster, thereby ensuring that all relevant information can be accurately extracted to form a complete structured text.

[0071] Although the K-means algorithm can perform feature clustering simply and directly, it is sensitive to noise and uneven distribution, which can easily cause noise data to pull the cluster center away, and long-tail fault labels to be missed or mislabeled. To improve the accuracy of label assignment, in another preferred embodiment, the present invention provides a method based on hierarchical clustering to achieve high-quality label classification and structured text conversion. The specific steps include:

[0072] Performing dimensionality reduction on the text feature vector, using a noise application space clustering algorithm to identify a core area of ​​the reduced text feature vector, and using a hierarchical density clustering algorithm to hierarchically divide the identified core area to obtain a plurality of subclusters;

[0073] Calculate the label similarity between each subcluster and each label field in the fixed label format, and determine the cluster label corresponding to each subcluster based on the label similarity. The label fields include project name, manufacturer, product type, product grade, problem type, problem location, problem cause, and specific problem situation.

[0074] According to the cluster label, each sub-cluster is filled into the corresponding label field to obtain a structured text corresponding to the converter transformer fault case.

[0075] In this embodiment, a dimensionality reduction algorithm is first used to reduce the dimensionality of the text feature vector, and the distance distortion of the high-dimensional space is avoided by retaining the local semantic structure. Preferably, the UMAP (Uniform Manifold Approximation and Projection) algorithm is used for feature dimensionality reduction. The UMAP algorithm achieves visual dimensionality reduction of data by minimizing the loss function. After the dimensionality reduction is completed, a noise application spatial clustering algorithm is used to automatically identify noise points and preliminarily divide dense areas. The noise points here refer to incorrectly labeled or abnormal samples. Preferably, the DBSCAN (Density-Based Spatial Clustering of Applications with Noise) clustering algorithm is used for preliminary division. For the core area divided by the DBSCAN density clustering algorithm, this embodiment further divides it by a hierarchical density clustering algorithm, thereby dynamically generating semantically consistent subclusters. Preferably, the HDBSCAN (Hierarchical Density-Based Spatial Clustering of Applications with Noise) algorithm is used to subdivide or merge the candidate clusters output by DBSCAN to obtain multiple subclusters.

[0076] The label similarity between each subcluster and each label field preset in the fixed label format is then calculated. In this embodiment, weighted cosine similarity is used to quantify the degree of semantic matching between the subcluster and the preset label field. When the label similarity between a subcluster and a label is greater than a first threshold, the label is used as the cluster label corresponding to the subcluster. The calculation formula for label similarity can be expressed as:

[0077]

[0078] In the formula, C represents the classification subcluster, which represents the text feature set to be matched; L k represents the keyword set of the preset k-th tag field. Each tag field corresponds to a set of keywords. For example, the keywords of the tag field "problem type" include "overheating" and "short circuit". w represents the keywords in subcluster C, that is, the subcluster keywords. Specifically, they are high-weight words selected by the TF-IDF algorithm. For example, w in a subcluster is "casing", "discharge", and "insulating material". Represents the keywords in the keyword set of the k-th label field, also known as the label keywords; TF-IDF(w) represents the TF-IDF value of sub-cluster keyword w in sub-cluster C, which is used to measure its importance within the sub-cluster. Its calculation formula can refer to the common calculation formula of the TF-IDF algorithm; Indicates tag keywords In the k-th class label field L k TF-IDF value in; ||C|| represents the vector modulus length of subcluster C; ||L k || represents the k-th label field L k The vector modulus of .

[0079] Label mapping is performed through label similarity to determine the labels corresponding to each sub-cluster. Then, the keywords of the sub-cluster can be filled into the corresponding label fields in the fixed label format to generate the structured text of the fault case.

[0080] When performing label mapping, if each subcluster can clearly match a single label, then the labels can be matched directly through the clustering results, thereby achieving the conversion of unstructured text to structured text. However, due to the complexity and diversity of fault cases, classification subclusters may be mapped to multiple labels, that is, multi-label coupled fault cases may occur. For example, subcluster C is simultaneously associated with "Problem Type - Overload" and "Problem Cause - Cooling Failure". For the case of multi-label coupling, one method is to directly perform label mapping, that is, by defining strong association rules, the high-frequency words appearing in the subcluster are strongly associated with the preset labels, and the labels are directly assigned instead of relying on similarity calculations. For example, in the strong association rules, keywords, corresponding labels, and keyword weights are set. When multiple keywords appear in a subcluster and are associated with different labels, the keyword with the highest weight is selected for label assignment.

[0081] Another approach is to optimize label allocation through conflict resolution based on node embedding. This approach can not only optimize label allocation for multi-label coupling, but also for situations where sub-cluster label similarity is too low. The specific optimization steps for conflict resolution include:

[0082] Determining, based on the label similarities, whether each subcluster satisfies a conflict resolution condition, wherein the conflict resolution condition includes that at least two label similarities are greater than a first threshold, or that all label similarities are less than a second threshold, and the first threshold is greater than the second threshold;

[0083] In response to the existence of a sub-cluster that meets the conflict resolution condition, the sub-cluster that meets the conflict resolution condition is used as a sub-cluster to be optimized;

[0084] A label co-occurrence graph is constructed using each label field in the fixed label format as a node and the label co-occurrence frequency in the historical fault case library as the edge weight.

[0085] Using a graph embedding algorithm to perform node embedding calculation on the tag co-occurrence graph to obtain a node embedding vector corresponding to each node in the tag co-occurrence graph;

[0086] Calculate the embedding similarity between the center vector of the subcluster to be optimized and the node embedding vector, and determine the cluster label corresponding to the subcluster to be optimized according to the embedding similarity.

[0087] In this embodiment, it is first determined whether there is multi-label coupling or label irrelevance. Multi-label coupling means that a subcluster has multiple labels with high similarity, that is, at least two label similarities between the subcluster and each label are greater than a first threshold, and the first threshold is preferably set to 85%. Label irrelevance means that none of the label similarities between the subcluster and each label is greater than a second threshold, and the second threshold is preferably set to 30%. If there is multi-label coupling or label irrelevance, it is determined that the subcluster meets the conflict resolution conditions. In this case, the subcluster is used as the subcluster to be optimized, and then conflict resolution is performed on the subcluster to be optimized to achieve label allocation optimization.

[0088] For subclusters to be optimized that require label allocation optimization, this embodiment adopts a conflict resolution method based on node embedding. Specifically, a label co-occurrence graph is first constructed based on a preset fixed label format and a historical fault case library. Each label field in the fixed label format is used as a graph node, that is, each node represents a label category. Each node is initialized as a TF-IDF vector. The TF-IDF value is calculated by the TF-IDF value of the label keyword in the historical fault case. Then, based on the historical fault case library, the co-occurrence relationship between the label fields is statistically analyzed. For example, "Manufacturer - TBEA" and "Problem Location - Winding" often appear at the same time. Therefore, the edge weight is defined by calculating the label co-occurrence frequency. The calculation formula for the edge weight is:

[0089]

[0090] Where, L i represents the i-th label, L j represents the jth label, W(L i , L j ) represents the edge weight between the i-th label and the j-th label, Count(L i ∩L j ) represents the number of times the i-th label and the j-th label appear together, Count(L i ) represents the number of times the i-th label appears independently, Count(L j ) represents the number of times the jth label appears independently.

[0091] In this embodiment, the edge type is an undirected edge, the edge weight range is [0, 1], and edges above the weight threshold are retained as valid associations. For the generated label co-occurrence graph, a graph embedding method is used to calculate the node embedding vector of the label co-occurrence graph. Preferably, the GraphSAGE (Graph Sample and Aggregated) model is used to sample neighborhood nodes of the label co-occurrence graph and aggregate features to generate the node embedding vector. The specific generation steps can refer to the conventional steps for calculating the node embedding vector using the GraphSAGE model, which will not be repeated here.

[0092] Then, the embedding similarity between the subcluster to be optimized and each label node in the label co-occurrence graph is calculated. In this embodiment, the cosine similarity between the center vector of the subcluster to be optimized and the node embedding vector of the label node is used as the embedding similarity. Then, the cluster label corresponding to the subcluster is determined based on the maximum value of each embedding similarity.

[0093] In fact, when cluster labels are selected by embedding similarity, it is still possible that two embedding similarities are close. This may be caused by blurred semantic boundaries or noise interference. In order to avoid this situation, in another preferred embodiment, this embodiment screens the optimal labels by setting a similarity difference threshold. Specifically, the difference between each embedding similarity and other embedding similarities is calculated. If the difference between the embedding similarity of a certain label and the other embedding similarities is greater than the similarity difference threshold, the label is assigned to the subcluster, thereby avoiding misselection due to minor differences.

[0094] If there is a judgment condition that does not meet the similarity difference threshold, it is considered that there is fierce competition between labels, and deep verification will be triggered at this time. Due to the two-level judgment of label similarity and embedding similarity, in actual applications, the probability of triggering deep verification is low. Therefore, for sub-clusters that require deep verification, the rationality of label allocation can be verified through manual review. Of course, the knowledge graph logic verification can also be performed through the pre-built knowledge graph of integrated power equipment to determine whether the label combination violates the knowledge base rules, and a logically reasonable label combination is selected based on the knowledge graph verification results. For example, "overload" should be associated with "problem type" rather than "manufacturer" first. The specific verification method can be flexibly selected according to the actual situation, and there are no excessive restrictions here.

[0095] This embodiment uses a label co-occurrence graph to model the implicit relationship between labels, optimizes the label mapping logic through graph embedding calculation, and prioritizes the allocation of high-frequency co-occurring label combinations, thereby resolving the label conflict problem. This embodiment improves the text quality of structured text by optimizing the label allocation of subclusters.

[0096] After generating structured text, text verification is required to ensure its quality. This verification process primarily includes consistency, accuracy, and completeness checks. This ensures that the generated structured text accurately reflects the characteristics of converter transformer fault cases and provides a reliable basis for subsequent analysis and decision-making. Specifically, consistency verification aims to confirm that the same characteristics are consistently expressed across different fault cases. For example, "low-end" and "low-end type" should be unified as "low-end" to avoid inconsistent descriptions across different samples. Secondly, accuracy verification evaluates the match between automatically generated labels and ground-truth labels by comparing them with expert-annotated data. Furthermore, completeness verification ensures that each fault case covers all feature categories and that no labels for any category are omitted. Combining manual verification with automated checks can more comprehensively identify potential issues and correct inconsistent, erroneous, or missing text, thereby ensuring high text data quality.

[0097] See also Figure 2 Based on the same inventive concept, a second embodiment of the present invention provides a structured processing system for converter transformer fault cases, including:

[0098] The data processing module 10 is used to clean and unify the format of the original text data of the converter transformer fault case to obtain text data;

[0099] A feature extraction module 20 is configured to pre-process the text data and input the pre-processed text data into a preset feature extraction model to obtain a text feature vector. The feature extraction model is constructed based on a natural language processing model. The pre-processing includes term segmentation and length unification.

[0100] The text generation module 30 is configured to classify the text feature vectors using a cluster analysis algorithm, and generate structured text corresponding to the converter transformer fault case based on the classification result and a preset fixed label format.

[0101] The technical features and effects of the structured processing system for converter transformer fault cases proposed in the embodiments of the present invention are the same as those of the method proposed in the embodiments of the present invention and are not further elaborated here. Each module in the structured processing system for converter transformer fault cases described above can be implemented in whole or in part through software, hardware, or a combination thereof. Each of these modules can be embedded in or independent of a processor in a computer device in hardware form, or stored in a computer device memory in software form, allowing the processor to call and execute the corresponding operations of each module.

[0102] In addition, an embodiment of the present invention further provides a computer device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the steps of the above method when executing the computer program.

[0103] See also Figure 3 , a diagram of the internal structure of a computer device in one embodiment, which may specifically be a terminal or a server. The computer device includes a processor, memory, a network interface, a display, and an input device connected via a system bus. The processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system and a computer program. The internal memory provides an environment for the operation of the operating system and computer program in the non-volatile storage medium. The network interface of the computer device is used to communicate with an external terminal via a network connection. When the computer program is executed by the processor, a structured processing method for converter transformer fault cases is implemented. The display screen of the computer device may be a liquid crystal display or an electronic ink display screen, and the input device of the computer device may be a touch layer covering the display screen, or may be a key, trackball, or touchpad provided on the computer device housing, or may be an external keyboard, touchpad, or mouse.

[0104] It can be understood by those skilled in the art that Figure 3 The structure shown in the figure is only a block diagram of a part of the structure related to the solution of the present application, and does not constitute a limitation on the computer device to which the solution of the present application is applied. The specific computing device may include more or fewer components than shown in the figure, or combine certain components, or have the same component arrangement.

[0105] In addition, an embodiment of the present invention further provides a computer-readable storage medium having a computer program stored thereon, which implements the steps of the above method when the computer program is executed by a processor.

[0106] In summary, the embodiments of the present invention propose a method, system, device, and medium for structured processing of converter transformer fault cases. The method cleans and unifies the format of the original text data of the converter transformer fault cases to obtain text data; preprocesses the text data, and inputs the preprocessed text data into a preset feature extraction model to obtain a text feature vector. The feature extraction model is constructed based on a natural language processing model, and the preprocessing includes term segmentation and length unification; the text feature vector is classified by a clustering analysis algorithm, and the structured text corresponding to the converter transformer fault case is generated according to the classification results and a preset fixed label format. The present invention extracts features from converter transformer fault cases through natural language processing technology, solves the problem of feature extraction caused by redundant text information of fault cases and diversified expressions of professional terms, and improves the completeness of feature extraction; improves the accuracy of label assignment through a label assignment method based on cluster analysis, and achieves high-quality structured text conversion, thereby providing accurate and efficient data support for intelligent analysis of converter transformer fault cases.

[0107] Each embodiment in this specification is described in a progressive manner, and the same or similar parts of each embodiment can be directly referred to each other, and each embodiment focuses on the differences from other embodiments. In particular, for the system embodiment, since it is basically similar to the method embodiment, the description is relatively simple, and the relevant parts can be referred to the partial description of the method embodiment. It should be noted that the various technical features of the above embodiments can be combined arbitrarily. In order to make the description concise, not all possible combinations of the various technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0108] The above-described embodiments merely represent several preferred implementations of the present application. While the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the patent. It should be noted that a person skilled in the art could make several improvements and substitutions without departing from the technical principles of the present invention, and these improvements and substitutions should also be considered within the scope of protection of the present application. Therefore, the scope of protection of the present patent application shall be based on the scope of protection of the claims.

Claims

1. A structured processing method for converter transformer fault cases, characterized in that: include: Clean and format the original text data of converter transformer fault cases to obtain text data; Preprocessing the text data and inputting the preprocessed text data into a preset feature extraction model to obtain a text feature vector, wherein the feature extraction model is constructed based on a natural language processing model, and the preprocessing includes term segmentation and length unification; The text feature vectors are classified by a cluster analysis algorithm, and structured text corresponding to the converter transformer fault case is generated according to the classification result and a preset fixed label format.

2. The structured processing method for converter transformer fault cases according to claim 1 is characterized in that: The step of cleaning and formatting the original text data of the converter transformer fault case to obtain the text data includes: Perform data cleaning on the original text data of converter transformer fault cases; According to preset modules, the cleaned original text data is divided into modules to obtain text data in a unified format. The preset modules include a problem description module, a cause analysis module, a treatment measure module and a treatment result module.

3. The structured processing method for converter transformer fault cases according to claim 1, characterized in that: The step of preprocessing the text data comprises: Dynamically segmenting professional terms in the electric power field in the text data using a byte pair encoding algorithm; The length of the initial text segments obtained by dynamic word segmentation is unified to obtain text segments.

4. The structured processing method for converter transformer fault cases according to claim 1 is characterized in that: The feature extraction model is built based on the RoBERTa model, the Adam optimizer is used to tune the model parameters, and the cross entropy loss is used as the loss function.

5. The structured processing method for converter transformer fault cases according to claim 1 is characterized in that: The step of classifying the text feature vectors by a cluster analysis algorithm and generating a structured text corresponding to the converter transformer fault case according to the classification result and a preset fixed label format includes: Classifying the text feature vectors using a K-means clustering algorithm to obtain a classification result; According to the classification result and a preset fixed label format, the text feature vector is filled into the corresponding label field to obtain a structured text of the converter transformer fault case; The tag fields include project name, manufacturer, product type, product grade, problem type, problem location, problem cause and specific problem situation.

6. The structured processing method for converter transformer fault cases according to claim 1, characterized in that: The step of classifying the text feature vectors by a cluster analysis algorithm and generating a structured text corresponding to the converter transformer fault case according to the classification result and a preset fixed label format includes: Performing dimensionality reduction on the text feature vector, using a noise application space clustering algorithm to identify a core area of ​​the reduced text feature vector, and using a hierarchical density clustering algorithm to hierarchically divide the identified core area to obtain a plurality of subclusters; Calculate the label similarity between each subcluster and each label field in the fixed label format, and determine the cluster label corresponding to each subcluster based on the label similarity. The label fields include project name, manufacturer, product type, product grade, problem type, problem location, problem cause, and specific problem situation. According to the cluster label, each sub-cluster is filled into the corresponding label field to obtain a structured text corresponding to the converter transformer fault case.

7. The structured processing method for converter transformer fault cases according to claim 6 is characterized in that: The step of determining the cluster labels corresponding to the subclusters according to the label similarity includes: Determining, based on the label similarities, whether each subcluster satisfies a conflict resolution condition, wherein the conflict resolution condition includes that at least two label similarities are greater than a first threshold, or that all label similarities are less than a second threshold, and the first threshold is greater than the second threshold; In response to the existence of a sub-cluster that meets the conflict resolution condition, the sub-cluster that meets the conflict resolution condition is used as a sub-cluster to be optimized; A label co-occurrence graph is constructed using each label field in the fixed label format as a node and the label co-occurrence frequency in the historical fault case library as the edge weight. Using a graph embedding algorithm to perform node embedding calculation on the tag co-occurrence graph to obtain a node embedding vector corresponding to each node in the tag co-occurrence graph; Calculate the embedding similarity between the center vector of the subcluster to be optimized and the node embedding vector, and determine the cluster label corresponding to the subcluster to be optimized according to the embedding similarity.

8. A structured processing system for converter transformer fault cases, characterized in that: include: A data processing module is used to clean and unify the format of the original text data of the converter transformer fault case to obtain text data; A feature extraction module is used to preprocess the text data and input the preprocessed text data into a preset feature extraction model to obtain a text feature vector, wherein the feature extraction model is constructed based on a natural language processing model, and the preprocessing includes term segmentation and length unification; The text generation module is used to classify the text feature vectors by using a cluster analysis algorithm, and generate structured text corresponding to the converter transformer fault case according to the classification result and a preset fixed label format.

9. A computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein: When the processor executes the computer program, the steps of the method according to any one of claims 1 to 7 are implemented.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 7 are implemented.

Images