Power grid operation and inspection standard intelligent matching method based on bidirectional dynamic adaptation

Abstract

The invention provides a power grid operation inspection standard intelligent matching method based on bidirectional dynamic adaptation, and the method comprises the steps: obtaining post names and post responsibility descriptions, generating post semantic vectors, constructing a post semantic knowledge base, and updating the post semantic vectors when the post responsibility descriptions are changed; obtaining a multi-source heterogeneous standard file, analyzing standard names, technical elements and revision historical information, constructing a structured standard database, and updating semantic vectors of standard terms during standard revision; based on post semantic vectors and standard information, a bidirectional dynamic adaptation mechanism of posts and standards is constructed by adopting a semantic vector matching and keyword association method, a matching relationship is updated when post responsibilities or standard terms are changed, and a matching result is stored to support standardized adaptation query and dynamic updating. According to the invention, accurate matching of post requirements and technical standards can be realized, the matching efficiency is improved, the cost of manual comparison is reduced, and the intelligent level of operation and inspection standard management is enhanced.

Description

Technical Field

[0001] The present invention belongs to the technical field of standardized dynamic management of power systems, and particularly relates to an intelligent matching method for grid operation and maintenance standards based on bidirectional dynamic adaptation. Background Art

[0002] During the process of grid operation and maintenance (operation and inspection), each position needs to strictly follow the technical standards, operation and maintenance regulations, and operation manuals formulated by the State Grid Corporation of China and the industry to ensure the safe operation of equipment and the standardization of operation and inspection work. However, with the upgrade of grid equipment types, the change of operating environment, and the continuous revision of standard documents, the matching relationship between job responsibilities and operation and inspection standards is also constantly adjusted. The existing standard matching methods mainly rely on manual search and comparison of standard clauses, which are difficult to meet the requirements of efficient and accurate matching in grid operation and inspection management.

[0003] In the prior art, Chinese Patent Application CN112612902A discloses a method and device for constructing a knowledge graph of main grid equipment, belonging to the technical field of grid operation and inspection. Through continuous exploration and experiments, the present invention structurally processes the basic information, operation data, and operation content of main grid equipment, and realizes the extraction of knowledge triples, ontology construction, completion, and reasoning of the knowledge graph of main grid equipment through system tools, thereby constructing a knowledge graph of main grid equipment. However, this patent mainly focuses on knowledge organization and reasoning at the equipment level, and has limited support for the matching of job responsibilities description and technical standards. First, the knowledge graph method mainly relies on structured entity relationship modeling, and it is difficult to effectively process complex natural language texts such as job responsibilities and standard clauses, resulting in limited matching accuracy. Second, the existing knowledge graph update mechanism relies on manually defined ontology structures and rule matching. Facing the frequent revision of operation and inspection standards, it is difficult to timely adjust the adaptation relationship between positions and standards. In addition, the knowledge graph mainly supports one-way queries, and it is difficult to achieve bidirectional dynamic adaptation between job requirements and standard clauses, and cannot meet the linkage matching requirements of job responsibility adjustment and standard revision. Summary of the Invention

[0004] The purpose of the present invention is to overcome the above-mentioned defects existing in the prior art and provide an intelligent matching method for grid operation and maintenance standards based on bidirectional dynamic adaptation.

[0005] The purpose of the present invention can be achieved by the following technical solutions:

[0006] On the one hand, the present invention provides a method for matching grid operation and maintenance standards based on bidirectional semantic dynamic adaptation, including the following steps:

[0007] Step S1: Obtain the job names and job responsibility descriptions in the "State Grid Corporation of China Power Supply Enterprise Job Classification Standard", generate job semantic vectors using the BERT-MLM-BiLSTM-Attention model, construct a job semantic knowledge base, and update the job semantic vectors when the job responsibility descriptions change;

[0008] Step S2: Obtain multi-source heterogeneous standard documents, parse the standard names, technical elements, and revision history information, construct a structured standard database, and update the semantic vectors of the standard clauses when the standards are revised;

[0009] Step S3: Based on the job semantic vectors and standard information, adopt the semantic vector matching and keyword association method to construct a two-way dynamic adaptation mechanism between jobs and standards, update the matching relationship when the job responsibilities or standard clauses change, and store the matching results to support standardized adaptation query and dynamic update.

[0010] Further, in the above-mentioned step S1, the construction of the job semantic knowledge base includes:

[0011] Extract the job names and their corresponding job responsibility descriptions in the "State Grid Corporation of China Power Supply Enterprise Job Classification Standard";

[0012] Generate job semantic vectors for the job names and job responsibility descriptions using the pre-trained BERT-MLM-BiLSTM-Attention model;

[0013] Use the MongoDB document database to store the job structured data, where the job structured data includes job names, job responsibility descriptions, job semantic vectors, and update times;

[0014] When the job responsibility description changes, use the difflib library in python to locate the added, deleted, or modified paragraphs, and only re-parse and vectorize the affected parts.

[0015] Further, the generation of job semantic vectors using the pre-trained BERT-MLM-BiLSTM-Attention model specifically includes:

[0016] Concatenate the job name and the job responsibility description into continuous text, add the [CLS] start marker and [SEP] separator marker through the BERT tokenizer, and uniformly truncate or pad to 512 tokens as the input to the BERT model;

[0017] Process the input text through the pre-trained BERT model, and obtain the hidden state matrix H output by the last layer Transformer of the pre-trained BERT model;

[0018] Taking the hidden state matrix H as the input, it is fed into the BiLSTM network to capture the temporal dependencies in the job description and generate a sequence of temporally enhanced vectors K = [k1, k2,..., k n , where k i is the vector of the i-th token output by the BiLSTM network;

[0019] The global attention mechanism is used to calculate the importance weights a i ;

[0020] Based on weighted summation, the final 768-dimensional job semantic vector V is calculated.

[0021] Furthermore, the weight assignment formula is:

[0022]

[0023] where a i is the attention weight of the i-th token vector, W k is a learnable attention weight matrix used to calculate the importance of each token, is the transpose of the i-th token vector output by the BiLSTM, and n is the total number of tokens in the sequence of temporally enhanced vectors generated by the BiLSTM.

[0024] Furthermore, the calculation formula for the job semantic vector V is:

[0025]

[0026] where V is the final 768-dimensional job semantic vector after dimensionality reduction, used for semantic representation of job responsibilities.

[0027] Furthermore, the training process of the BERT-MLM-BiLSTM-Attention model is as follows:

[0028] Based on the text data in the power grid operation and maintenance field, a special corpus for power grid operation and maintenance is constructed by collecting "State Grid Corporation of China's Power Supply Enterprise Job Classification Standard", power operation and maintenance regulations, equipment operation manuals, fault reports, maintenance operation guides, etc. The corpus is subjected to data cleaning, duplicate removal, and format standardization processing, and power operation and maintenance terms and job responsibility descriptions are extracted to form a term list;

[0029] Based on the above-mentioned power grid operation and maintenance specific corpus, use the Masked Language Model task to train the BERT model, mask the power grid operation and maintenance terms with a 15% probability to generate masked texts, and use the BERT model to predict the masked terms. Compare the predicted words with the original terms, calculate the cross-entropy loss, optimize the BERT model parameters, and obtain the trained BERT model applicable to the semantic analysis of job responsibilities;

[0030] Based on the trained BERT model and the loss function, train the BiLSTM network and the learnable attention weight matrix W k for training.

[0031] Furthermore, the training of the BiLSTM network and the learnable attention weight matrix W based on the trained BERT model and the loss function k includes the following specific steps:

[0032] Based on the trained BERT model, perform vectorization processing on the names and job responsibility descriptions of different positions to obtain the hidden state matrix of the job responsibility descriptions;

[0033] Pass the hidden state matrix of the job responsibility descriptions through the BiLSTM network and the learnable attention weight matrix W k for processing to obtain the job semantic vectors of different positions, and distinguish positions based on the job category label C, where C represents the category to which the position belongs, and different job categories have different C values;

[0034] According to the obtained job semantic vectors of different positions and their corresponding job category labels C, calculate the loss function, and use Mini-Batch gradient descent combined with the AdamW optimizer to adjust the parameters of the BiLSTM network and the attention weight matrix.

[0035] Furthermore, the loss function is as follows:

[0036]

[0037] where is the loss function, N is the number of training samples, V i and V j represent the job semantic vectors of position i and position j respectively, d(V i , V j ) = ∥V i - V j ∥2 is the Euclidean distance of the job semantic vectors, y i,j is the job similarity label. When position i and position j are the same, y i,j = 1, otherwise y i,j = 0, m is the distance threshold, and C iand C i represent the category labels of position i and position j respectively, and λ is a hyperparameter for weighing position category information.

[0038] Furthermore, the standard information acquisition process includes:

[0039] Based on multi-source heterogeneous standard documents, collect the operation and maintenance regulations, equipment technical standards, operation guides, safety management regulations of State Grid Corporation and relevant industry standards, and construct a standard document database;

[0040] Perform formatting processing on the standard document database, extract standard names, standard numbers, release times, scopes of application, technical requirements, key terms and revision history information by using text parsing technology, and store them as structured standard data;

[0041] Based on the structured standard data, construct a standard information index library, classify the standard terms hierarchically by using rule matching and natural language processing technology, generate standard category labels, and establish standard revision relationships according to standard version information;

[0042] When the standard document is revised, based on the standard version information, compare the content of the old and new standard terms, use the text difference detection method to locate the modified part, and update the semantic vector of the standard terms based on semantic similarity calculation to achieve dynamic update of standard information.

[0043] Furthermore, the specific steps of step S3 include:

[0044] Based on the position semantic vector and standard information, construct a two-way dynamic adaptation mechanism between the position and the standard, perform semantic vector matching and keyword association analysis on the position responsibility description and standard term content, and obtain a preliminary matching result;

[0045] Calculate the similarity of the preliminary matching results, calculate the matching score based on the Euclidean distance between the position semantic vector and the standard term vector, and perform classification constraints in combination with the position category label and the standard category label to screen out the matching results with high relevance;

[0046] When the position responsibility description changes, based on the position semantic knowledge base, extract the position semantic vectors before and after the change, calculate the cosine similarity of the vectors before and after the change, and update the adaptation relationship between the position and the standard according to the similarity change;

[0047] When the standard terms are revised, based on the standard version information, extract the standard semantic vectors before and after the revision, calculate the Euclidean distance of the semantic vectors, and combine keyword matching to analyze the change range of the revised content, perform re-adaptation calculation on the affected positions, and update the matching relationship between the positions and the standards;

[0048] Based on the final matching results of job responsibilities and standard clauses, construct the mapping relationship between job requirements and technical specifications, and store it in the job standard matching database to support the standardized adaptation query and dynamic update of job requirements.

[0049] Compared with the prior art, the present invention has the following advantages:

[0050] (1) Through the job semantic vector modeling technology based on the BERT-MLM-BiLSTM-Attention model, the present invention realizes the deep semantic representation of job responsibility descriptions, can accurately capture the context semantic information of job texts, and improves the accuracy of job responsibility semantic expression compared with traditional keyword matching or static word vector methods, making the matching between jobs and operation and maintenance standards more accurate.

[0051] (2) Through the structured parsing and version management technology based on multi-source heterogeneous standard files, the present invention realizes the automatic extraction, classified storage and revision history tracing of standard clauses, can quickly identify the changed content when the standard files are updated, and synchronously adjust the matching job responsibilities. Compared with the traditional manual comparison of standard clauses, it improves the timeliness and accuracy of job and standard adaptation, and ensures that job responsibilities are consistent with the latest operation and maintenance standards.

[0052] (3) Through the two-way dynamic adaptation mechanism based on job semantic vectors and standard information, the present invention realizes the two-way matching between job responsibilities and standard clauses, so that not only can applicable standard clauses be retrieved according to job requirements, but also the matching job responsibilities can be automatically adjusted after the standards are revised, avoiding the problem of mismatch between job requirements and standard requirements caused by the traditional one-way matching mode, and improving the flexibility and adaptability of job standard management.

[0053] (4) Through the intelligent matching algorithm based on semantic vector matching combined with keyword association, the present invention realizes the accurate association between job responsibilities and technical standards, can dynamically adjust the matching relationship when job descriptions change or standards are revised, reduces human intervention compared with relying on manual experience or rule matching, improves the degree of intelligence and automation of matching, and reduces the maintenance cost of job and standard adaptation. Brief Description of the Drawings

[0054] Figure 1 is the flowchart of the method of the present invention;

[0055] Figure 2 is the radial tree diagram of power operation and maintenance related jobs based on the "Power Supply Enterprise Job Classification Standard of State Grid Corporation of China" of the present invention. Detailed Embodiment

[0056] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, rather than all embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the scope of protection of the present invention.

[0057] Embodiment 1:

[0058] This embodiment provides a power grid operation and maintenance standard matching method based on bidirectional semantic dynamic adaptation, as Figure 1 shown, including the following steps:

[0059] Step S1: Obtain the job names and job responsibility descriptions in the "Power Supply Enterprise Job Classification Standard of State Grid Corporation of China", use the BERT-MLM-BiLSTM-Attention model to generate job semantic vectors, construct a job semantic knowledge base, and update the job semantic vectors when the job responsibility descriptions change;

[0060] Step S2: Obtain multi-source heterogeneous standard files, parse the standard names, technical elements, and revision history information, construct a structured standard database, and update the semantic vectors of the standard clauses when the standards are revised;

[0061] Step S3: Based on the job semantic vectors and standard information, adopt a semantic vector matching and keyword association method to construct a two-way dynamic adaptation mechanism between the job and the standard, update the matching relationship when the job responsibilities or standard clauses change, and store the matching results to support standardized adaptation query and dynamic update.

[0062] Further, in the step S1, the construction of the semantic vector library includes:

[0063] Extract the job names and their corresponding job responsibility descriptions in the "Power Supply Enterprise Job Classification Standard of State Grid Corporation of China";

[0064] Among them, as Figure 2 shown, the radial tree diagram of the operation and maintenance field-related jobs selected from the "Power Supply Enterprise Job Classification Standard of State Grid Corporation of China" covers the job middle class names and job subclass names; among them, there is a general responsibility description for each middle class job, explaining the main functions of this type of job in the entire power grid technology system. The specific work tasks that each subclass job needs to undertake are listed in detail in the subclass job description.

[0065] For example, the middle class of transmission operation and maintenance technology is a technical job that ensures the safety, reliability, and efficiency of the power grid system and is responsible for the operation and maintenance of transmission equipment.

[0066] This middle class is divided into three sub - classes: a) Transmission line operation and maintenance technology, b) Transmission cable operation and maintenance technology, c) Transmission equipment condition assessment technology.

[0067] The main work of position c) Transmission equipment condition assessment technology includes:

[0068] 1) AC and DC transmission equipment condition detection and evaluation;

[0069] 2) Operation and maintenance of the main station system for equipment condition detection, management of equipment condition information, and practical evaluation of the production informatization system;

[0070] 3) Equipment condition assessment, fault diagnosis, professional technical analysis, and technical appraisal of retired equipment;

[0071] 4) Sampling inspection of power grid equipment performance, handover tests, diagnostic tests, live detection, as well as on - line monitoring, live detection equipment, electrical test instruments and meters for network access detection and regular verification;

[0072] 5) Whole - process technical supervision of the maintenance profession, equipment manufacturing supervision, technical exchanges and training, etc., and supervision of the implementation of technical standards and anti - accident measures

[0073] Generate job semantic vectors for the job name and job responsibility description using the pre - trained BERT - MLM - BiLSTM - Attention model;

[0074] Use the MongoDB document database to store job structured data, where the job structured data includes job name, job responsibility description, job semantic vector, and update time;

[0075] When the job responsibility description changes, use the difflib library in python to locate the added, deleted, or modified paragraphs, and only re - parse and vectorize the affected parts.

[0076] Furthermore, the generation of job semantic vectors using the pre - trained BERT - MLM - BiLSTM - Attention model specifically includes:

[0077] Concatenate the job name and job responsibility description into continuous text, add the [CLS] start marker and [SEP] separator marker through the BERT tokenizer, and uniformly truncate or pad to 512 tokens as the input to the BERT model;

[0078] Process the input text through the pre - trained BERT model, and obtain the hidden state matrix H output by the last layer Transformer of the pre - trained BERT model;

[0079] Taking the hidden state matrix H as the input, it is fed into the BiLSTM network to capture the temporal dependencies in the job description and generate a sequence of temporally enhanced vectors K = [k1, k2,..., k n , where k i is the vector of the i-th token output by the BiLSTM network;

[0080] The global attention mechanism is used to calculate the importance weights a i ;

[0081] Based on weighted summation, the final 768-dimensional job semantic vector V is calculated.

[0082] The present invention uses a pre-trained BERT-MLM-BiLSTM-Attention model to generate job semantic vectors, so as to achieve a deep semantic understanding of job names and job responsibility descriptions, making the vector representation of job descriptions more semantically consistent and discriminative.

[0083] First, the job name and the job responsibility description are concatenated into continuous text, and [CLS] start marker and [SEP] end marker are added through the BERT tokenizer. At the same time, it is uniformly truncated or padded to 512 tokens to meet the input format requirements of the BERT model. The purpose of doing this is to ensure that the BERT model can process the input text completely and avoid affecting the learning effect of the model due to inconsistent text lengths. In addition, through [CLS], the BERT model generates a global semantic representation, and [SEP] is used as a separator for different text segments, enabling the model to better understand the relationship between the job name and the responsibility description.

[0084] Subsequently, the pre-trained BERT model is used to process the input text, and the hidden state matrix H output by the last layer of Transformer of BERT is obtained. The self-attention mechanism of BERT can capture the long-distance dependencies between different words in the sentence, making the context semantic information of the job responsibility description more complete. However, the BERT model itself mainly focuses on word-level semantic expressions and is difficult to model the temporal dependencies of sequence information. Therefore, it is necessary to further combine with the BiLSTM network.

[0085] Taking the hidden state matrix H as the input, it is fed into the BiLSTM network to perform bidirectional encoding on the job description text, thereby capturing the temporal dependencies in the job description and obtaining a sequence of temporally enhanced vectors. BiLSTM can comprehensively consider the front and back information of the text and is particularly suitable for dealing with the sequential dependencies involved in job responsibility descriptions. For example, although the words in "responsible for inspecting equipment" and "the person in charge of equipment inspection" are the same, but the order is different, and BiLSTM can recognize their semantic differences.

[0086] To further improve the quality of semantic vectors, the present invention uses a global attention mechanism to calculate the importance weights of each token vector in the job description. The introduction of the attention mechanism enables the model to automatically learn the key information in the job description and assigns higher weights to the key tokens. For example, in a description like "Responsible for substation equipment inspection and fault troubleshooting", "inspection" and "fault troubleshooting" are the core tasks, while "responsible for" is a general expression. The attention mechanism can effectively allocate weights to make the job semantic vector better highlight the core responsibilities of the job.

[0087] Finally, based on weighted summation, the final 768-dimensional job semantic vector V is calculated to ensure that the most critical semantic features of the job description information can still be retained during the dimensionality reduction process. The semantic vector generated in this way can not only accurately represent the job responsibilities but also be used for efficient matching of operation and maintenance standards, improve the adaptability between the job and the standards, and at the same time reduce the interference of redundant information on the matching results, making the job semantic representation more compact and having better discrimination ability.

[0088] Furthermore, the weight assignment formula is:

[0089]

[0090] where a i is the attention weight of the i-th token vector, W k is a learnable attention weight matrix used to calculate the importance of each token, is the transpose of the i-th token vector output by the BiLSTM, and n is the total number of tokens in the sequence of time series enhanced vectors generated by the BiLSTM.

[0091] The present invention uses the attention mechanism to calculate the importance weights of each token to ensure that during the generation of the semantic vector of the job description, the key information that has a greater impact on the core responsibilities of the job can be highlighted, while reducing the interference of irrelevant or general words on the semantic vector.

[0092] Furthermore, the calculation formula for the job semantic vector V is:

[0093]

[0094] where V is the final 768-dimensional job semantic vector after dimensionality reduction, used for semantic representation of job responsibilities.

[0095] Furthermore, the training process of the BERT-MLM-BiLSTM-Attention model is:

[0096] Based on the text data in the field of power grid operation and maintenance inspection, a special corpus for power grid operation and maintenance inspection is constructed by collecting the "Power Supply Enterprise Post Classification Standard of State Grid Corporation of China", power operation and maintenance regulations, equipment operation manuals, fault reports, maintenance operation guides, etc. The data in the corpus is cleaned, duplicate entries are removed, and the format is standardized. Power operation and maintenance terms and job responsibility descriptions are extracted to form a term glossary;

[0097] Based on the above-mentioned special corpus for power grid operation and maintenance inspection, the BERT model is trained using the Masked Language Model task. The power grid operation and maintenance terms are masked with a probability of 15%, and masked texts are generated. The masked terms are predicted by the BERT model, the predicted words are compared with the original terms, the cross-entropy loss is calculated, and the parameters of the BERT model are optimized to obtain the trained BERT model applicable to job responsibility semantic analysis;

[0098] Based on the trained BERT model and the loss function, the BiLSTM network and the learnable attention weight matrix W k are trained.

[0099] Furthermore, the training of the BiLSTM network and the learnable attention weight matrix W based on the trained BERT model and the loss function k specifically includes:

[0100] Based on the trained BERT model, the names and job responsibility descriptions of different positions in the input are vectorized to obtain the hidden state matrix of the job responsibility description;

[0101] The hidden state matrix of the job responsibility description is processed through the BiLSTM network and the learnable attention weight matrix W k to obtain the job semantic vectors of different positions, and job differentiation is performed based on the job category label C, where C represents the category to which the position belongs, and different job categories have different C values;

[0102] According to the obtained job semantic vectors of different positions and their corresponding job category labels C, the loss function is calculated, and the parameters of the BiLSTM network and the attention weight matrix are adjusted using Mini-Batch gradient descent combined with the AdamW optimizer.

[0103] The training process of the BERT-MLM-BiLSTM-Attention model of the present invention aims to construct a semantic vector model that can accurately represent job responsibility descriptions to achieve efficient matching between positions and standard terms. First, text data in the power grid operation and maintenance field is collected to construct a special corpus for power grid operation and maintenance, and data cleaning, deduplication, and formatting processing are carried out. Power operation and maintenance terms and job responsibility descriptions are extracted to form a term vocabulary. The purpose of this step is to ensure the quality of the model training data, eliminate redundant information, improve the standardization degree of job descriptions, and enable subsequent semantic modeling to more accurately capture the core content of job responsibilities.

[0104] Based on the corpus, the BERT model is trained using the Masked Language Model (MLM) task. The power grid operation and maintenance terms are masked with a probability of 15%, and the BERT model is used to predict the masked terms. The cross-entropy loss between the prediction result and the original term is calculated, and the BERT model parameters are optimized. This training strategy can adaptively fine-tune the BERT model in the power grid operation and maintenance field, enabling it to have stronger industry-specific semantic understanding ability when processing job descriptions, and improving the professionalism and accuracy of semantic vectors.

[0105] Based on the trained BERT model, the BiLSTM network and the learnable attention weight matrix are trained to further enhance the quality of job semantic vectors. Specifically, first, the trained BERT model is used to vectorize the names and job responsibility descriptions of different positions to obtain the hidden state matrix of job descriptions. Since the BERT model mainly captures word-level semantic information, and job descriptions usually have strong temporal dependencies, the hidden state matrix needs to be input into the BiLSTM network to model the temporal information of job descriptions, thereby generating a sequence of temporally enhanced semantic vectors.

[0106] Next, the learnable attention weight matrix is used to calculate the attention weights of each token to highlight the core information in the job description and obtain the final job semantic vector. To ensure that the job semantic vector can not only reflect the details of job responsibilities but also maintain a reasonable class discrimination degree, the model introduces the job category label C for constraint, where different job categories have different C values. This category constraint can make the job semantic vectors of the same category closer, while the job vectors of different categories are relatively dispersed, improving the hierarchical expression ability of job descriptions.

[0107] To optimize the training process of the BiLSTM network and the attention weight matrix, a loss function is used to calculate the similarity of job semantic vectors and perform contrastive learning in combination with job category information. The loss function enhances the consistency of the semantic expressions of the same job by minimizing the Euclidean distance of the job semantic vectors of the same job, and at the same time maximizes the Euclidean distance between the job semantic vectors of different jobs to ensure the separability between jobs. In addition, the job category label CC further constrains the distribution of the semantic vectors, making the projections of jobs in the same category more hierarchical in the high-dimensional space and improving the accuracy of job classification and matching.

[0108] During the training process, Mini-Batch gradient descent combined with the AdamW optimizer is used to adjust the parameters of the BiLSTM network and the attention weight matrix. The AdamW optimizer can dynamically adjust the learning rate and combine the weight decay mechanism to effectively prevent overfitting and improve the generalization ability of the model. Mini-Batch gradient descent can reduce the computational overhead, accelerate the training convergence speed, and at the same time perform gradient updates on small batches of data, enabling the model to more stably learn the semantic features of job descriptions.

[0109] Through the above training process, the present invention realizes a high-quality semantic representation of job descriptions, enabling the job semantic vectors to accurately capture the core content of job responsibilities and effectively distinguish different categories of jobs, thereby improving the matching accuracy between jobs and standards. Compared with traditional keyword matching or static word vector models, this method has stronger semantic understanding ability, can adapt to different expression forms of job descriptions, and improves the intelligent adaptation level of job requirements and operation and maintenance standards.

[0110] Furthermore, the loss function is:

[0111]

[0112] where is the loss function, N is the number of training samples, V i and V j represent the job semantic vectors of job i and job j respectively, d(V i ,V j ) = ∥V i - V j ∥2 is the Euclidean distance of the job semantic vectors, y i,j is the job similarity label. When job i and job j are the same, y i,j = 1, otherwise y i,j = 0, m is the distance threshold, C i and C i represent the category labels of job i and job j respectively, and λ is a hyperparameter for weighing job category information.

[0113] Further, the standard information acquisition process includes:

[0114] Based on multi-source heterogeneous standard documents, collect the operation and maintenance regulations, equipment technical standards, operation guides, safety management regulations of State Grid Corporation and relevant industry standards, and construct a standard document database;

[0115] Perform formatting processing on the standard document database, extract standard names, standard numbers, release times, scopes of application, technical requirements, key terms and revision history information by using text parsing technology, and store them as structured standard data;

[0116] Based on the structured standard data, construct a standard information index library, use rule matching and natural language processing technologies to classify and layer standard terms, generate standard category labels, and establish standard revision relationships according to standard version information;

[0117] When the standard document is revised, based on the standard version information, compare the content of the old and new standard terms, use the text difference detection method to locate the modified part, and update the semantic vector of the standard terms based on the semantic similarity calculation to achieve the dynamic update of standard information.

[0118] Further, the specific steps of step S3 include:

[0119] Based on the job semantic vector and standard information, construct a two-way dynamic adaptation mechanism between the job and the standard, perform semantic vector matching and keyword correlation analysis on the job responsibility description and standard term content, and obtain a preliminary matching result;

[0120] Calculate the similarity of the preliminary matching result, calculate the matching score based on the Euclidean distance between the job semantic vector and the standard term vector, and perform classification constraints in combination with the job category label and the standard category label to screen out the matching results with high relevance;

[0121] When the job responsibility description changes, based on the job semantic knowledge base, extract the job semantic vectors before and after the change, calculate the cosine similarity of the vectors before and after the change, and update the adaptation relationship between the job and the standard according to the similarity change;

[0122] When the standard terms are revised, based on the standard version information, extract the standard semantic vectors before and after the revision, calculate the Euclidean distance of the semantic vectors, and combine keyword matching to analyze the scope of change of the revised content, perform re-adaptation calculation on the affected jobs, and update the matching relationship between the job and the standard;

[0123] Based on the final matching result of the job responsibility and the standard terms, construct a mapping relationship between the job requirements and technical specifications, and store it in the job-standard matching database to support the standardized adaptation query and dynamic update of job requirements.

[0124] Example 2:

[0125] The parts not mentioned in this example are the same as those in Example 1.

[0126] This example provides an electronic device, including a memory and a processor. A computer program is stored on the memory, and when the processor executes the program, it implements a power grid operation and maintenance standard matching method based on bidirectional semantic dynamic adaptation as described in any one of the above.

[0127] This example provides a computer-readable storage medium, on which a computer program is stored. When the program is executed by a processor, it implements a power grid operation and maintenance standard matching method based on bidirectional semantic dynamic adaptation as described in any one of the above.

[0128] If the above functions are implemented in the form of software function units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or a part of this technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions for causing a computer device (which may be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the methods described in various embodiments of the present invention. The foregoing storage medium includes: USB flash drives, mobile hard disks, read-only memories (ROMs), random access memories (RAMs), magnetic disks, or optical discs and other various media that can store program codes.

[0129] As described above, the above are only specific embodiments of the present invention, but the protection scope of the present invention is not limited thereto. Any person skilled in the art within the technical scope disclosed by the present invention can easily think of various equivalent modifications or substitutions, and these modifications or substitutions should all be covered by the protection scope of the present invention. Therefore, the protection scope of the present invention should be subject to the protection scope of the claims.

Claims

1. A smart matching method for power grid operation and maintenance standards based on bidirectional dynamic adaptation, characterized in that, It includes the following steps: Step S1: Obtain the job names and job responsibility descriptions in the "Power Supply Enterprise Job Classification Standard of State Grid Corporation of China", generate job semantic vectors using the BERT-MLM-BiLSTM-Attention model, construct a job semantic knowledge base, and update the job semantic vectors when the job responsibility descriptions change; Step S2: Obtain multi-source heterogeneous standard documents, parse the standard names, technical elements, and revision history information, construct a structured standard database, and update the semantic vectors of the standard clauses when the standards are revised; Step S3: Based on the job semantic vectors and standard information, adopt the semantic vector matching and keyword association method to construct a two-way dynamic adaptation mechanism between jobs and standards, update the matching relationship when the job responsibilities or standard clauses change, and store the matching results to support standardized adaptation query and dynamic update.

2. The intelligent matching method for power grid operation and maintenance standards based on bidirectional dynamic adaptation according to claim 1, characterized in that In the above Step S1, the construction of the job semantic knowledge base includes: Extract the job names and their corresponding job responsibility descriptions in the "Power Supply Enterprise Job Classification Standard of State Grid Corporation of China"; Generate job semantic vectors for the job names and job responsibility descriptions using the pre-trained BERT-MLM-BiLSTM-Attention model; Use the MongoDB document database to store the job structured data, where the job structured data includes job names, job responsibility descriptions, job semantic vectors, and update times; When the job responsibility descriptions change, use the difflib library in python to locate the added, deleted, or modified paragraphs, and only re-parse and vectorize the affected parts.

3. The intelligent matching method for grid operation and maintenance standards based on two-way dynamic adaptation according to claim 2, characterized in that The generation of the job semantic vectors using the pre-trained BERT-MLM-BiLSTM-Attention model specifically includes: Concatenate the job name and job responsibility description into continuous text, add the [CLS] start marker and [SEP] separator marker through the BERT tokenizer, and uniformly truncate or pad to 512 tokens as the input to the BERT model; Process the input text through the pre-trained BERT model, and obtain the hidden state matrix H output by the last layer Transformer of the pre-trained BERT model; Taking the hidden state matrix H as the input, it is fed into the BiLSTM network to capture the temporal dependencies in the job description and generate a sequence of temporally enhanced vectors K = [k1, k2,..., k n , where k i is the vector of the i-th token output by the BiLSTM network; Use the global attention mechanism to calculate the importance weight a of each token vector in the job description i ; Calculate the final 768-dimensional job semantic vector V based on weighted summation.

4. A method for intelligent matching of power grid operation and maintenance standards based on bidirectional dynamic adaptation according to claim 3, characterized in that, The weight assignment formula is: where a i is the attention weight of the i-th token vector, and W k is a learnable attention weight matrix used to calculate the importance of each token, is the transpose of the i-th token vector output by the BiLSTM, and n is the total number of tokens in the sequence of time series enhanced vectors generated by the BiLSTM.

5. The intelligent matching method for grid operation and maintenance standards based on bidirectional dynamic adaptation according to claim 3, characterized in that, The calculation formula for the job semantic vector V is: where V is the final 768-dimensional job semantic vector after dimensionality reduction, used for the semantic representation of job responsibilities.

6. The intelligent matching method for power grid operation and maintenance standards based on bidirectional dynamic adaptation according to claim 3, wherein, The training process of the BERT-MLM-BiLSTM-Attention model is: Based on the text data in the power grid operation and maintenance field, collect the "Power Supply Enterprise Job Classification Standard of State Grid Corporation of China", power grid operation and maintenance regulations, equipment operation manuals, fault reports, maintenance operation guides, etc. to construct a special corpus for power grid operation and maintenance, perform data cleaning, deduplication, and format standardization processing on the corpus, extract power grid operation and maintenance terms and job responsibility descriptions, and form a term list; Based on the above-mentioned power grid operation and maintenance specific corpus, the BERT model is trained using the Masked Language Model task. The power grid operation and maintenance terms are masked with a probability of 15% to generate masked text, and the masked terms are predicted by the BERT model. The cross-entropy loss is calculated by comparing the predicted words with the original terms, and the parameters of the BERT model are optimized to obtain the trained BERT model suitable for semantic analysis of job responsibilities. Based on the trained BERT model and loss function, train the BiLSTM network and the learnable attention weight matrix W k for training.

7. A method for intelligent matching of power grid operation and maintenance standards based on two-way dynamic adaptation according to claim 6, characterized in that, Training the BiLSTM network and the learnable attention weight matrix W based on the trained BERT model and the loss function k Specifically, the training process includes: Based on the trained BERT model, the names and job responsibility descriptions of different positions in the input are vectorized to obtain the hidden state matrix of the job responsibility description. The hidden state matrix of the job responsibility description is processed through a BiLSTM network and a learnable attention weight matrix W k to obtain job semantic vectors for different positions, and job differentiation is performed based on the job category label C, where C represents the category to which the position belongs, and different job categories have different C values; According to the obtained job semantic vectors of different positions and their corresponding job category labels C, the loss function is calculated, and the parameters of the BiLSTM network and the attention weight matrix are adjusted using Mini-Batch gradient descent combined with the AdamW optimizer.

8. A method for intelligent matching of power grid operation and maintenance standards based on bidirectional dynamic adaptation according to claim 7, characterized in that, The loss function is as follows: Among them, is the loss function, N is the number of training samples, V i and V j respectively represent the job semantic vectors of job i and job j. d(V i , V j ) = ∥V i - V j ∥² is the Euclidean distance of the job semantic vectors. y i,j is the job similarity label. When job i and job j are the same, y i,j = 1; otherwise, y i,j = 0. m is the distance threshold. C i and C i respectively represent the category labels of job i and job j. λ is the hyperparameter for weighing job category information.

9. The intelligent matching method for grid operation and maintenance standards based on two-way dynamic adaptation according to claim 1, characterized in that The process of obtaining the standard information includes: Based on multi-source heterogeneous standard documents, the operation and maintenance regulations, equipment technical standards, operation guides, safety management regulations of the State Grid Corporation and relevant industry standards are collected to construct a standard document database. The standard document database is formatted, and text parsing technology is used to extract the standard name, standard number, release time, scope of application, technical requirements, key clauses and revision history information, and store them as structured standard data. Based on the structured standard data, a standard information index library is constructed. The standard clauses are hierarchically classified using rule matching and natural language processing technology to generate standard category labels, and the standard revision relationship is established according to the standard version information. When the standard document is revised, based on the standard version information, the content of the new and old standard clauses is compared, the modified part is located using the text difference detection method, and the semantic vector of the standard clause is updated based on the semantic similarity calculation to realize the dynamic update of the standard information.

10. A method for intelligent matching of power grid operation and maintenance standards based on two-way dynamic adaptation according to claim 1, characterized in that, The specific steps of step S3 include: Based on the job semantic vector and the standard information, a two-way dynamic adaptation mechanism between the job and the standard is constructed, and semantic vector matching and keyword association analysis are performed on the job responsibility description and the standard clause content to obtain a preliminary matching result. The similarity of the preliminary matching result is calculated, the matching score is calculated based on the Euclidean distance between the job semantic vector and the standard clause vector, and classification constraints are combined with the job category label and the standard category label to screen the matching results with high relevance. When the job responsibility description changes, based on the job semantic knowledge base, the job semantic vectors before and after the change are extracted, the cosine similarity of the vectors before and after the change is calculated, and the adaptation relationship between the job and the standard is updated according to the similarity change. When the standard clause is revised, based on the standard version information, the standard semantic vectors before and after the revision are extracted, the Euclidean distance of the semantic vectors is calculated, and the change range of the revised content is analyzed by combining keyword matching. The affected jobs are re-adapted and calculated, and the matching relationship between the job and the standard is updated. Based on the final matching result of the job responsibility and the standard clause, a mapping relationship between the job requirements and the technical specifications is constructed and stored in the job-standard matching database to support the standardized adaptation query and dynamic update of the job requirements.

Images