Cross-batch power data intelligent labeling method

Abstract

The invention discloses a cross-batch electric power data intelligent labeling method, which comprises the following steps of: constructing a hierarchical label scene tree, scientifically dividing data category levels according to an electric power safety supervision business process, formulating a unified label standard, integrating industry specifications and actual demands, and defining label meanings, ranges and values; automatic standard tools such as basic rules and machine learning are combined with interactive labeling methods such as active learning and crowdsourcing labeling to improve labeling efficiency; a Word2Vec word vector model and a semantic similarity calculation, dynamic updating and consistency detection method are adopted to realize semantic efficient mapping, and the method has the capabilities of quickly adjusting a label system and adaptively optimizing an annotation and mapping algorithm for a new data type and format, improves the cross-batch data processing capability of electric power safety supervision, and improves the data processing efficiency of the electric power safety supervision. The defects in the aspects of label consistency, labeling efficiency, semantic mapping accuracy and new data adaptability in the prior art are overcome.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence data annotation, and in particular to a method for intelligent annotation of cross-batch power data. Background Technology

[0002] In the field of artificial intelligence data annotation, existing technologies face numerous challenges in intelligent annotation and semantic mapping across batches of data. First, label consistency is difficult to guarantee. Currently, data from different batches originates from diverse sources, including monitoring equipment installed at different times and data from different manufacturers. Due to the lack of unified and standardized labeling criteria and effective management methods, label definitions for different batches of data are often vague and inconsistent. For example, when annotating the operating status of power equipment, differences exist in the labeling of data from different operators or different time periods, hindering subsequent data integration and analysis, and making it impossible to accurately perform fault diagnosis and risk assessment based on labels. Second, annotation efficiency is low. Traditional data annotation relies heavily on manual annotation, a process that is not only time-consuming and labor-intensive, but also far too slow to keep up with the speed of data generation when dealing with massive amounts of power data. Furthermore, manual annotation is susceptible to subjective factors; fatigue and differences in professional skills among annotators can lead to errors. For instance, when annotating a large amount of power equipment parameter data, manual annotation may result in mislabeling and omissions, and its efficiency is extremely low, severely impacting the timeliness of power safety monitoring data processing. Third, semantic mapping effects are unsatisfactory. Existing semantic mapping techniques often rely on simple dictionary matching or shallow machine learning models, failing to fully grasp the complex semantic relationships within the power sector. Power industry terminology carries specific contextual meanings, and new terms constantly emerge with technological advancements. Current technologies struggle to handle these complexities, leading to inaccurate semantic mapping and hindering in-depth data mining and analysis of power data. For instance, when interpreting power equipment fault reports, semantic mapping errors may result in misjudgments of the fault's cause, impacting subsequent maintenance decisions and the stable operation of the power system.

[0003] Therefore, it is necessary to conduct research on intelligent annotation and semantic mapping technologies for cross-batch data in the field of power safety supervision. By establishing unified labeling standards, introducing automated annotation methods, and employing semantic mapping algorithms, this research aims to solve the problems of inconsistent labels, low annotation efficiency, and inaccurate semantic mapping in existing technologies. This will improve the intelligence level of power safety supervision, ensure the safe and stable operation of the power system, and meet the growing data processing needs of the power industry. Currently, there are numerous related patents in the field of data annotation, covering multiple aspects from image data annotation to multimodal data annotation and natural language processing.

[0004] Most existing intelligent data annotation technologies focus on specific domains or types of data annotation, and have limitations in practical applications. For example, while some methods theoretically possess high accuracy and efficiency, they may require significant computational resources and time in practice, limiting their application in large-scale data annotation tasks. Furthermore, some methods are not well-suited to the diversity and complexity of data, failing to adequately handle cross-domain or cross-modal data annotation needs, ultimately resulting in less than ideal outcomes. Summary of the Invention

[0005] The purpose of this invention is to solve the problems in the prior art.

[0006] The technical solution adopted by this invention to solve its technical problem is: to provide a method for intelligent labeling of cross-batch power data, comprising the following steps:

[0007] Based on the physical structure and operation process of the power system and the actual needs of power safety supervision, a tag scene tree with a clear hierarchical structure is constructed.

[0008] Establish a unified labeling standard that covers the label definition, value range, applicable conditions, labeling rules, relationship with other labels, and update mechanism;

[0009] Based on the aforementioned labeling standard, a multi-method collaborative annotation method is constructed;

[0010] A semantic similarity-based mapping method enables intelligent labeling of cross-batch power data.

[0011] Preferably, the tag scene tree starts from the top level, and each node i has rich attributes. It includes the node name, description, parent node number, node type, associated metadata, and applicable annotation rule set; for node i and its parent node j, not only is there a parent-child relationship P(i)=j, but also the inheritance relationship between nodes is defined, that is, the child node inherits some attributes and annotation rules of the parent node, and can be extended or refined according to its own characteristics.

[0012] Preferably, the construction of the label scene tree with a clear hierarchical structure adopts a top-down recursive algorithm. First, the top-level node and its attributes are determined, and then the nodes of the next level are generated step by step according to the classification knowledge of the power system and business logic.

[0013] For each node, the existing node structure is traversed using either Depth-First Search (DFS) or Breadth-First Search (BFS) algorithms to ensure that the newly generated node is in a reasonable position and has an accurate relationship with other nodes.

[0014] During node generation, the node attributes are initialized and assigned values ​​according to pre-defined rules and constraints, and a corresponding set of annotation rules is set.

[0015] Preferably, the method for constructing a multi-method collaborative annotation system includes:

[0016] Automated labeling, with preset basic rules, enables rapid preprocessing of large-scale data through a classification model based on these basic rules;

[0017] Interactive annotation involves actively learning annotation methods to select the most valuable samples for manual annotation in complex scenes and with fuzzy features.

[0018] Preferably, the automated annotation method for constructing a multi-method collaborative annotation system includes the following steps:

[0019] Establish clear basic rules for the power sector, and directly generate annotation results based on these basic rules;

[0020] By reusing historical annotation experience and training a classification model using already labeled power safety monitoring data, the model learns the mapping relationship between "text / image features and labels" to achieve automatic annotation of new data.

[0021] We augment the sample with unlabeled data and guide the labeling of unlabeled data with a small amount of labeled data through semi-supervised learning.

[0022] Preferably, the solidified power sector defines fundamental rules, which include:

[0023] Keyword matching rules are based on an electrical terminology dictionary to accurately match keywords in text or images and map them to corresponding tags;

[0024] Regular expression rules are used to extract and annotate formatted structured information using regular expressions.

[0025] Logical judgment rules are designed based on the conditional logic of power operation specifications.

[0026] Preferably, the step of augmenting samples with unlabeled data and guiding the labeling of unlabeled data with a small amount of labeled data through semi-supervised learning includes the following steps:

[0027] Feature construction transforms text data from a high-dimensional, sparse symbolic representation into a low-dimensional, dense numerical vector representation, i.e., embedding vectors.

[0028] Similarity calculation uses cosine similarity to calculate the similarity between feature vectors of each pair of data points;

[0029] Label propagation begins by initializing a label probability matrix Y(n, x, c), where n is the total number of samples and c is the number of categories. For labeled data, Y_i is the one-hot encoding of its true label. For unlabeled data, Y_i is initialized to [0,0,0] or a uniform distribution. Then, a transition probability matrix T is constructed, where T_{ij} represents the probability of transitioning from node j to node i. Finally, iterative updates are performed, with the label probability of each node updated in each iteration as a weighted average of the label probabilities of all its neighboring nodes. After the iterations, each unlabeled node U_i obtains a stable label probability vector P_i=[p_i1,p_i2,...,p_ic], where p_ic represents the confidence that the node belongs to category c.

[0030] The results were screened, and a confidence threshold was set. Samples with a confidence level greater than the confidence threshold were added to the training dataset; samples with a confidence level less than or equal to the confidence threshold were sent to the active learning loop or directly labeled by domain experts.

[0031] Preferably, the interactive annotation includes the following steps:

[0032] An active learning system architecture is constructed, consisting of a sample selection module, a labeling request module, a model update module, and a labeling feedback module. The sample selection module selects representative samples from a large number of unlabeled samples based on an uncertainty sampling strategy. The labeling request module sends the selected samples to the labelers and provides detailed labeling guidelines and relevant reference information. After the labelers complete the labeling, they return the labeling results to the system through the labeling feedback module. The model update module uses the newly labeled data to update the classification model, realizing continuous optimization of the model.

[0033] Develop an uncertainty sampling strategy, comprehensively consider all strategies to calculate a comprehensive uncertainty score for each unlabeled sample, and select image samples with large entropy values ​​based on the scores to send to the labelers.

[0034] The newly labeled data undergoes quality checks. After data augmentation, the qualified data is added to the training set to retrain the classification model. Incremental learning is used during training to avoid forgetting previously learned knowledge.

[0035] Preferably, the uncertainty sampling strategy includes an entropy-based sampling strategy, a marginal sampling strategy, and a committee query-based sampling strategy;

[0036] The entropy-based sampling strategy measures uncertainty by calculating the entropy of the model's predicted probability distribution for unlabeled samples; the higher the entropy, the higher the uncertainty.

[0037] The marginal sampling strategy selects the sample with the smallest difference between the two highest probability categories predicted by the model, because the smaller the difference, the less uncertain the model is about classifying the sample.

[0038] The committee-based query sampling strategy trains multiple different models and selects the sample with the greatest difference in prediction results among these models.

[0039] Preferably, the semantic similarity-based mapping method for intelligent labeling of cross-batch power data includes:

[0040] Data preparation; comprehensively collect equipment operation and maintenance logs, operation process manuals, safety regulations and systems within the power industry, as well as multi-source information such as technical reports released by industry authorities, academic research results from research institutions, and publicly available news reports on power safety accidents; perform format standardization and cleaning; and carry out in-depth data cleaning and processing.

[0041] Text preprocessing; text segmentation; stop word filtering based on the constructed stop word list; selective part-of-speech tagging and stemming operations;

[0042] Word vector conversion; inputting power industry terminology and related text data into the Word2Vec model for training; after training, each power industry term is mapped to a low-dimensional dense word vector;

[0043] Calculate similarity; measure semantic similarity between words by calculating the cosine similarity between word vectors;

[0044] Model application; using semantic similarity for annotation expansion and semantic understanding.

[0045] The present invention has the following beneficial effects:

[0046] (1) This invention has significant advantages over existing technologies through innovations in label consistency, labeling efficiency, and semantic processing capabilities. It brings higher efficiency, more accurate decision-making basis and stronger technical adaptability to power safety supervision work, and powerfully promotes the intelligent development of the power industry.

[0047] (2) This invention fundamentally solves the problem of the lack of unified standards for data labeling by constructing a hierarchical label scene tree and a unified label standard. The hierarchical label scene tree provides a clear organizational structure for various data labels. From the macroscopic power system architecture to the microscopic equipment details, each can find its exact location in the tree structure and the relationship between them is clear. The unified label standard specifies in detail the labeling specifications, terminology definitions, etc. for different data types. This enables that in the field of power safety supervision, regardless of the equipment used to collect the data or the stage of the project, the labeling can be strictly carried out in accordance with the unified standard.

[0048] (3) This invention utilizes automated standard tools such as basic rules, machine learning, and semi-supervised learning to customize and develop data tailored to the characteristics of power safety monitoring data, enabling rapid processing of large amounts of routine data annotation tasks. Machine learning and semi-supervised learning algorithms can automatically learn annotation patterns from historical data and quickly annotate new data, reducing the workload of manual annotation. At the same time, the active learning annotation method uses intelligent algorithms to select the most representative and uncertain data samples, which are then handed over to the annotation personnel for processing, making the annotation work more targeted. This approach leverages the accuracy advantages of manual annotation while reducing unnecessary manual annotation workload.

[0049] (4) The semantic similarity-based mapping algorithm developed in this invention adopts the Word2Vec word vector model and the consistency detection method of semantic similarity and dynamic update mechanism, which improves the semantic processing capability.

[0050] The present invention will be further described in detail below with reference to the accompanying drawings and embodiments, but the present invention is not limited to the embodiments. Attached Figure Description

[0051] Figure 1 This is a diagram illustrating the method steps of an embodiment of the present invention;

[0052] Figure 2 This is a flowchart illustrating an embodiment of the present invention;

[0053] Figure 3 This is a schematic diagram of the power safety monitoring image tagging system according to an embodiment of the present invention;

[0054] Figure 4 This is a schematic diagram of the unlabeled data processing flow according to an embodiment of the present invention. Detailed Implementation

[0055] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this invention, and not all of them. Based on the embodiments of this invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this invention.

[0056] See Figure 1 and Figure 2 The diagram shown is a flowchart and step diagram of a method according to an embodiment of the present invention, including the following steps:

[0057] S101, based on the physical structure and operation process of the power system and the actual needs of power safety supervision, constructs a tag scene tree with a clear hierarchical structure;

[0058] S102, to establish a unified labeling standard, covering the label definition, value range, applicable conditions, labeling rules, relationship with other labels, and update mechanism;

[0059] S103, Based on the aforementioned labeling standard, construct a multi-method collaborative annotation method;

[0060] S104, a semantic similarity-based mapping method for intelligent labeling of cross-batch power data.

[0061] Specifically, in S101, a tag scene tree with a clear hierarchical structure is constructed based on the physical structure and operational process of the power system, as well as the actual needs of power safety monitoring. (See also...) Figure 3 As shown, the top level of this tree represents the overall classification of power safety monitoring images, which can be divided into primary categories such as equipment, environment, and operation. Taking equipment as an example, it is further subdivided into secondary categories such as power generation equipment, transmission equipment, substation equipment, and distribution equipment. For transmission equipment, it can be further subdivided into tertiary categories such as transmission lines, towers, and insulators, and so on, delving into appropriate levels according to the actual situation to ensure comprehensive and detailed coverage of all types of objects that may appear in power safety monitoring images.

[0062] As shown in Table 1, the hierarchical label scene tree starts from the top level, and each node i has rich attributes. In addition to node name, description, and parent node number, the annotation also includes node type (such as equipment class, environment class, etc.), associated metadata (such as the range of technical parameters of the equipment, monitoring indicators of the environment, etc.), and the applicable set of annotation rules. For node i and its parent node j, not only is there a parent-child relationship P(i)=j, but an inheritance relationship between nodes is also defined. That is, the child node inherits some attributes and annotation rules from the parent node, and can extend or refine them according to its own characteristics. For example, the "pole tower" node inherits the basic attributes and general annotation rules related to power transmission from the "transmission equipment" node, while also possessing its own specific attributes and annotation rules regarding pole tower structure, height, etc.

[0063] Table 1: Tag Scene Tree Node Attributes

[0064]

[0065] A top-down recursive algorithm is used to construct the tag scene tree. First, the top-level nodes and their attributes are determined. Then, based on the classification knowledge of the power system and business logic, nodes at the next level are generated progressively. For each node, the existing node structure is traversed using either Depth-First Search (DFS) or Breadth-First Search (BFS) algorithms to ensure that the newly generated node is positioned appropriately and its relationship with other nodes is accurate. During node generation, the node attributes are initialized and assigned values ​​according to pre-defined rules and constraints. For example, when generating the "Transformer" node, based on the function and characteristics of transformers in the power system, its node description is determined as "Electrical equipment used to change the magnitude of AC voltage," and associated metadata includes rated capacity, voltage ratio, etc., and a corresponding set of labeling rules is set, such as rules for determining the model based on the transformer's appearance characteristics.

[0066] Specifically, in S102, a unified labeling standard is established. First, a standard system framework is established. A unified labeling standard is established, covering multiple aspects such as label definition, value range, applicable conditions, labeling rules, relationship with other labels, and update mechanism, forming a complete standard system framework, as shown in Table 2.

[0067] Table 2. Example of definition of core elements of label standards:

[0068]

[0069] Secondly, the labeling rules are refined. For each label, detailed and strict labeling rules are formulated. Taking the "insulator" label as an example, the labeling rules include: First, extracting the contour features of the insulator using image recognition technology. For example, the shape of an insulator is usually disc-shaped or rod-shaped. The geometric parameters of the contour (such as radius, height, number of discs, etc.) are calculated and matched with a standard insulator model. Simultaneously, the color features of the insulator are extracted. A normal insulator should be white or light gray. If the color changes significantly (such as yellowing or blackening), it must be noted in the label. Furthermore, the texture features of the insulator must be considered. The surface texture of a normal insulator should be smooth and uniform. If cracks, damage, or other texture changes occur, the fault type must be accurately labeled. Through multi-dimensional feature extraction and strict matching rules, the accuracy and consistency of the labeling are ensured.

[0070] Finally, standard document management and updates. The unified labeling standards are compiled into a standard document. This document not only contains detailed information on all the aforementioned labels but also provides example images and explanations to facilitate understanding and adherence by labelers. A version management mechanism for the standard document is established, and it is updated regularly as power technology develops, new equipment is introduced, and business needs change. The update process follows a strict approval process, requiring joint review by power industry experts, safety supervisors, and the labeling team to ensure the standard's scientific validity, practicality, and authority.

[0071] Specifically, the automated annotation in S103 achieves rapid preprocessing of large-scale data and reduces the workload of manual annotation by "solidifying domain knowledge with basic rules, reusing historical annotations through machine learning, and utilizing unlabeled data through semi-supervised learning". Its core is to replace repetitive labor with machines, leaving only complex scenarios to humans.

[0072] First, solidify clear rules within the power sector. For scenarios in power safety supervision where the "terminology-label" correspondence is clear (such as equipment name and operation type), labeling results are directly generated through preset rules, requiring no manual intervention. The basic rule design logic is as follows:

[0073] (1) Keyword matching rules: Based on the power terminology dictionary, keywords in text or images are accurately matched and mapped to corresponding tags. For example, if “SF6 circuit breaker” appears in the text, it is automatically labeled “Equipment type - high voltage circuit breaker”; if the appearance of “disconnecting switch” appears in the image, it is automatically labeled “Equipment name - disconnecting switch”.

[0074] (2) Regular expression rules: For structured information with format (such as equipment number, voltage level), use regular expressions to extract and label it. For example, use r'(\d{1,3})kV' to match "220kV" and "110kV", and automatically label "parameter-voltage level-XXkV".

[0075] (3) Logical judgment rules: Design conditional logic in conjunction with power operation specifications. For example, if "power outage" and "grounding switch closed" appear simultaneously in the text, the system will automatically label it as "operation type - safety measure execution".

[0076] Secondly, historical annotation experience is reused. The classification model is trained using already annotated power safety monitoring data, allowing the model to learn the mapping relationship between "text / image features and labels", and then automatically annotate new data.

[0077] Finally, the sample is expanded using unlabeled data. Addressing the current situation in the power safety monitoring field where there is "few labeled data and many unlabeled data," semi-supervised learning is used to guide the labeling of unlabeled data with a small amount of labeled data, reducing reliance on manual labeling. The core method is label propagation. See [link to relevant documentation]. Figure 4 As shown, it includes the following steps:

[0078] (1) Feature construction. The text data is transformed from a high-dimensional, sparse symbolic representation to a low-dimensional, dense numerical vector representation (i.e., embedding), so that mathematical similarity calculation can be performed.

[0079] (2) Similarity Calculation. Cosine similarity is used to calculate the similarity between feature vectors of each pair of data points. Cosine similarity focuses on the direction of vectors rather than their magnitude, making it very suitable for text similarity calculation. The formula is as follows:

[0080]

[0081] (3) Label propagation. First, initialize a label probability matrix Y(nxc), where n is the total number of samples and c is the number of categories. For labeled data, Y_i is the one-hot encoding of its true label (e.g., [0,1,0] for the "overload" class); for unlabeled data, Y_i is initialized to [0,0,0] or uniformly distributed. Then, construct a transition probability matrix T, where T_{ij} represents the probability of transitioning from node j to node i, usually obtained by normalizing the similarity matrix W. Finally, perform iterative updates. In each iteration, the label probability of each node is updated to the weighted average of the label probabilities of all its neighboring nodes: Y^{(t+1)}=T*Y^{(t)}. After the iteration, each unlabeled node U_i will obtain a stable label probability vector P_i=[p_i1, p_i2,...,p_ic], where p_ic represents the confidence that the node belongs to category c.

[0082] (4) Result screening. Set a confidence threshold, such as 0.8. For samples with a confidence score > 0.8, directly add them (data, pseudo-label) to the training dataset; for samples with a confidence score ≤ 0.8, send these samples into the active learning loop or directly hand them over to domain experts for labeling.

[0083] Specifically, interactive annotation in S103. While automated annotation is efficient, it is prone to errors in complex scenes and with ambiguous features. Therefore, it is necessary to actively learn annotation methods to select the most valuable samples for annotation, avoiding wasting human effort on low-value data, thereby improving annotation efficiency and model performance. This includes the following steps:

[0084] S1031, Build the active learning system architecture. The active learning system consists of a sample selection module, a labeling request module, a model update module, and a labeling feedback module. The sample selection module selects representative samples from a large number of unlabeled samples based on an uncertainty sampling strategy. The labeling request module sends the selected samples to the labelers and provides detailed labeling guidelines and relevant reference information. After the labelers complete the labeling, they return the labeling results to the system through the labeling feedback module. The model update module uses the newly labeled data to update the labeled model, realizing continuous model optimization.

[0085] S1032, Develop an uncertainty sampling strategy. Employ a combination of multiple uncertainty sampling strategies to improve the effectiveness of sample selection. The main strategies include entropy-based sampling, marginal sampling, and committee query-based sampling. Taking these strategies into account, calculate a comprehensive uncertainty score for each unlabeled sample, and select image samples with high entropy values ​​(i.e., high uncertainty) based on the score to send to the labelers.

[0086] Specifically, the uncertainty sampling strategy includes:

[0087] (1) Entropy-based sampling strategy: The uncertainty is measured by calculating the entropy of the model's predicted probability distribution for unlabeled samples. The higher the entropy, the higher the uncertainty. The formula is:

[0088]

[0089] Where P(y|x) is the probability that the model predicts sample x as class y.

[0090] (2) Based on the marginal sampling strategy: select the sample with the smallest difference between the two highest probability categories predicted by the model, because the smaller the difference, the more uncertain the model is about classifying the sample.

[0091] (3) Based on the committee query sampling strategy: train multiple different models, such as by different initialization parameters or data augmentation methods, and select the sample with the greatest difference in the prediction results of these models.

[0092] S1033, Model Update Module. This module performs quality checks on newly labeled data, such as verifying that the labels conform to a unified labeling standard and identifying any contradictions or errors. After data augmentation, the qualified data is added to the training set. During model retraining, incremental learning is employed to prevent the forgetting of previously learned knowledge. Specifically, different weights are assigned to old and new data during training, allowing the model to retain the features of the old data while learning from the new data. Simultaneously, the hyperparameters are fine-tuned based on the characteristics of the newly labeled data to further optimize model performance.

[0093] Specifically, the semantic similarity-based mapping algorithm implementation in S104 includes:

[0094] S1041, Data Preparation. First, comprehensively collect multi-source information such as equipment operation and maintenance logs, operation process manuals, safety regulations within the power industry, as well as technical reports authoritatively released by the industry, academic research results of scientific research institutions, and public news reports on power safety accidents. These data cover all-round semantic information in the field of power safety supervision, from daily operation specifications to complex accident analysis, providing a rich and representative data foundation for subsequent model training. Second, perform format unification and cleaning on the collected data. For data in different formats, such as PDF, DOC, XML, etc., use data conversion tools and scripts to losslessly convert them into plain text formats convenient for processing. With the help of technologies such as regular expressions and OCR recognition, remove noise content such as garbled characters, error characters, headers, footers, page numbers, etc. from the data. For data with complex structures such as tables and pictures, extract valuable text information and perform reasonable format adjustment. At the same time, use NER recognition technology to identify and extract key entities in the text, such as power equipment names, personnel positions, time, etc., and perform standardization processing on them. For example, unify different expressions of "transformer" into standard terms. In addition, restore abbreviations and shorthands in the text, such as restoring "kV" to "kilovolt", to enhance the readability and consistency of the text. Finally, perform in-depth data cleaning and sorting operations. Appropriately streamline semantic duplicates or redundant content in the text through text summarization technology, and retain the core semantic information. At the same time, establish a vocabulary normalization table to uniformly process variant words, simplified and traditional Chinese conversions, etc. in Chinese texts, such as standardizing "probability" to "probability" and converting "traditional Chinese characters" to "traditional Chinese characters", further improving the consistency and standardization of the text and creating favorable conditions for subsequent model training. See Table 3 for the comparison of vocabulary before and after processing.

[0095] Table 3 Example of Vocabulary Normalization:

[0096]

[0097] S1042, Text Preprocessing. First, select the Jieba word segmentation tool and combine it with a custom dictionary to perform word segmentation on the text. Jieba word segmentation can segment continuous text into basic word or phrase units according to the grammar and semantic rules of the Chinese language. To adapt to professional terms in the field of power safety supervision, construct a custom dictionary and include specific professional vocabulary, abbreviations, etc. in the industry. At the same time, for some compound words with special meanings, by adjusting the parameters or rules of Jieba word segmentation, ensure that it can accurately identify and reasonably segment. See Table 4 for the word segmentation results.

[0098] Table 4 Example of Word Segmentation Results:

[0099]

[0100] Secondly, stop word filtering is performed based on the constructed stop word list. This list not only includes common, general stop words but is also optimized for the power safety supervision field. A Bloom filter is used to identify and remove stop words from the text. Simultaneously, stop words are flexibly filtered using part-of-speech tagging and contextual analysis to avoid mistakenly deleting important information.

[0101] Finally, part-of-speech tagging and stemming are selectively performed. For part-of-speech tagging, a deep learning-based model is used, specifically a CRF-based model customized for the power sector. This model is based on a general CRF and fine-tuned using a pre-annotated power safety supervision corpus to enable the model to recognize domain-specific parts of speech. The tagging results are optimized using a Conditional Random Field (CRF), with the following probability formula:

[0102]

[0103] in, This is a normalization factor that ensures the sum of the probabilities of all possible part-of-speech sequences is 1. Characteristic function The weights are learned from the training data. The characteristic function describes the "previous word property". Current part of speech The relationship between “current text x” and “position i”.

[0104] For stemming, in English text processing, the Porter Stemmer is used to restore inflected words to their stems, eliminating the impact of tense and singular / plural differences on semantic matching. For example, "maintaining" becomes "maintain", and "failures" becomes "failure". In Chinese text processing, a standardized vocabulary table is established for unified processing.

[0105] S1043, Word2Vec word vector model. First, model training and optimization. Power industry terminology and related text data are input for training the Word2Vec model. During training, hyperparameters such as window size, vector dimension, and negative sampling quantity are fine-tuned to adapt to the characteristics of power industry text. Negative sampling technology is used to accelerate the model training process and improve training efficiency. To improve the model's generalization ability, the training data is preprocessed, including stop word removal, stemming, and part-of-speech tagging. Simultaneously, domain-specific corpora are introduced for expanded training, such as power industry standard documents and professional papers.

[0106] Secondly, semantic similarity is calculated. After training, each technical term in the power industry is mapped to a low-dimensional, dense word vector. The semantic similarity between word vectors is measured by calculating the cosine similarity, using the following formula:

[0107]

[0108] in, and Let represent the vector representations of the two words respectively, and · denote the vector dot product. This represents the norm of the vector. The closer the similarity value is to 1, the more semantically similar the two words are.

[0109] S1044, Model Application. In image annotation, semantic similarity is used for annotation expansion and semantic understanding. For example, when annotating a device in an image as a "disconnect switch," the semantic similarity between "disconnect switch" and other related terms (such as "knife switch") is calculated. If the similarity is high, "knife switch" can be added as a synonym to the annotation information, enriching the annotation content. Simultaneously, when processing user queries or data analysis, semantic similarity is used to match user-input keywords with already annotated terms, improving query accuracy and recall.

[0110] This invention patent ensures label consistency across different batches of data by constructing a hierarchical label scene tree and a unified label standard, providing a reliable foundation for data integration and analysis. It utilizes automated standard tools such as basic rules, machine learning, and semi-supervised learning, as well as active learning labeling methods, to improve labeling efficiency and rapidly process massive amounts of data. A semantic similarity-based mapping algorithm is developed to achieve efficient semantic mapping and accurately handle complex semantic relationships in the power sector. Simultaneously, the entire technical solution possesses good flexibility and scalability, enabling rapid adaptation to new data types and formats, and promoting the intelligent and efficient development of power safety supervision.

[0111] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for intelligent labeling of cross-batch power data, characterized in that, Includes the following steps: Based on the physical structure and operation process of the power system and the actual needs of power safety supervision, a tag scene tree with a clear hierarchical structure is constructed. Establish a unified labeling standard that covers the label definition, value range, applicable conditions, labeling rules, relationship with other labels, and update mechanism; Based on the aforementioned labeling standard, a multi-method collaborative annotation method is constructed; A semantic similarity-based mapping method enables intelligent labeling of cross-batch power data.

2. The intelligent labeling method for cross-batch power data according to claim 1, characterized in that, The label scene tree starts from the top level, and each node i has rich attributes. It includes the node name, description, parent node number, node type, associated metadata, and applicable annotation rule set; for node i and its parent node j, not only is there a parent-child relationship P(i)=j, but also the inheritance relationship between nodes is defined, that is, the child node inherits some attributes and annotation rules of the parent node, and can be extended or refined according to its own characteristics.

3. The intelligent labeling method for cross-batch power data according to claim 1, characterized in that, The construction of the label scene tree with a clear hierarchical structure adopts a top-down recursive algorithm. First, the top-level node and its attributes are determined, and then the nodes of the next level are generated step by step according to the classification knowledge of the power system and business logic. For each node, the existing node structure is traversed using either Depth-First Search (DFS) or Breadth-First Search (BFS) algorithms to ensure that the newly generated node is in a reasonable position and has an accurate relationship with other nodes. During node generation, the node attributes are initialized and assigned values ​​according to pre-defined rules and constraints, and a corresponding set of annotation rules is set.

4. The intelligent labeling method for cross-batch power data according to claim 1, characterized in that, The method for constructing a multi-method collaborative annotation system includes: Automated labeling, with preset basic rules, enables rapid preprocessing of large-scale data through a classification model based on these basic rules; Interactive annotation involves actively learning annotation methods to select the most valuable samples for manual annotation in complex scenes and with fuzzy features.

5. The intelligent labeling method for cross-batch power data according to claim 4, characterized in that, The method for constructing a multi-method collaborative annotation system, the automated annotation, includes the following steps: Establish clear basic rules for the power sector, and directly generate annotation results based on these basic rules; By reusing historical annotation experience, the classification model is trained using already labeled power safety monitoring data, allowing the model to learn the mapping relationship between "text / image features and labels" and achieve automatic annotation of new data; We augment the sample with unlabeled data and guide the labeling of unlabeled data with a small amount of labeled data through semi-supervised learning.

6. The intelligent labeling method for cross-batch power data according to claim 5, characterized in that, The established basic rules in the field of solidified power include: Keyword matching rules are based on an electrical terminology dictionary to accurately match keywords in text or images and map them to corresponding tags; Regular expression rules are used to extract and annotate formatted structured information using regular expressions. Logical judgment rules are designed based on the conditional logic of power operation specifications.

7. The intelligent labeling method for cross-batch power data according to claim 5, characterized in that, The method of augmenting samples with unlabeled data and guiding the labeling of unlabeled data with a small amount of labeled data through semi-supervised learning includes the following steps: Feature construction transforms text data from a high-dimensional, sparse symbolic representation into a low-dimensional, dense numerical vector representation, i.e., embedding vectors. Similarity calculation uses cosine similarity to calculate the similarity between feature vectors of each pair of data points; Label propagation begins by initializing a label probability matrix Y(n, x, c), where n is the total number of samples and c is the number of categories. For labeled data, Y_i is the one-hot encoding of its true label. For unlabeled data, Y_i is initialized to [0,0,0] or a uniform distribution. Then, a transition probability matrix T is constructed, where T_{ij} represents the probability of transitioning from node j to node i. Finally, iterative updates are performed, with the label probability of each node updated in each iteration as a weighted average of the label probabilities of all its neighboring nodes. After the iterations, each unlabeled node U_i obtains a stable label probability vector P_i=[p_i1, p_i2,...,p_ic], where p_ic represents the confidence that the node belongs to category c. Results screening involves setting a confidence threshold. Samples with a confidence level greater than the threshold are added to the training dataset; samples with a confidence level less than or equal to the threshold are sent to the active learning loop or directly labeled by domain experts.

8. The intelligent labeling method for cross-batch power data according to claim 4, characterized in that, The interactive annotation includes the following steps: An active learning system architecture is constructed, consisting of a sample selection module, a labeling request module, a model update module, and a labeling feedback module. The sample selection module selects representative samples from a large number of unlabeled samples based on an uncertainty sampling strategy. The labeling request module sends the selected samples to the labelers and provides detailed labeling guidelines and relevant reference information. After the labelers complete the labeling, they return the labeling results to the system through the labeling feedback module. The model update module uses the newly labeled data to update the classification model, realizing continuous optimization of the model. Develop an uncertainty sampling strategy, comprehensively consider all strategies to calculate a comprehensive uncertainty score for each unlabeled sample, and select image samples with large entropy values ​​based on the scores to send to the labelers. The newly labeled data undergoes quality checks. After data augmentation, the qualified data is added to the training set to retrain the classification model. Incremental learning is used during training to avoid forgetting previously learned knowledge.

9. The intelligent labeling method for cross-batch power data according to claim 8, characterized in that, The uncertainty sampling strategies include entropy-based sampling strategies, marginal sampling strategies, and committee query-based sampling strategies. The entropy-based sampling strategy measures uncertainty by calculating the entropy of the model's predicted probability distribution for unlabeled samples; the higher the entropy, the higher the uncertainty. The marginal sampling strategy selects the sample with the smallest difference between the two highest probability categories predicted by the model, because the smaller the difference, the less uncertain the model is about classifying the sample. The committee-based query sampling strategy trains multiple different models and selects the sample with the greatest difference in prediction results among these models.

10. The intelligent labeling method for cross-batch power data according to claim 1, characterized in that, The semantic similarity-based mapping method enables intelligent annotation of cross-batch power data, including: Data preparation; comprehensively collect equipment operation and maintenance logs, operation process manuals, safety regulations and systems within the power industry, as well as multi-source information such as technical reports released by industry authorities, academic research results from research institutions, and publicly available news reports on power safety accidents; perform format standardization and cleaning; and carry out in-depth data cleaning and processing. Text preprocessing; word segmentation of the text; stop word filtering based on the constructed stop word list; selective part-of-speech tagging and stemming operations; Word vector conversion; inputting power industry terminology and related text data into the Word2Vec model for training; after training, each power industry term is mapped to a low-dimensional dense word vector; Calculate similarity; measure semantic similarity between words by calculating the cosine similarity between word vectors; Model application; using semantic similarity for annotation expansion and semantic understanding.

Images