Relay protection intelligent processing method and device, electronic equipment and storage medium

Abstract

The invention relates to the technical field of intelligent processing of relay protection, and discloses an intelligent processing method and device for relay protection, electronic equipment and a storage medium, and the method comprises the steps: carrying out the preprocessing of a to-be-recognized text, so as to obtain a standardized text; inputting the standardized text into a pre-training language model for semantic coding to obtain a text coding vector; performing semantic similarity calculation on the text coding vector and a pre-stored event vector in a preset standard event graph; selecting a pre-stored event vector with the highest similarity as a target event vector according to a calculation result; obtaining a business flow chart corresponding to the target event vector; and performing automatic processing based on the business flow chart. The method has the beneficial effects that the real-time performance, the safety and the stability of a relay protection system are remarkably improved, the manual review burden is reduced, and the overall decision support capability is enhanced.

Description

Technical Field

[0001] This invention relates to the field of intelligent processing technology for relay protection, and in particular to an intelligent processing method, device, electronic device, and storage medium for relay protection. Background Technology

[0002] In ensuring the stable operation of power systems, relay protection devices undertake the core tasks of fault detection, isolation control, and rapid response. Especially in dispatch automation and maintenance, accurate and timely identification and response to alarm events are crucial for ensuring equipment safety and reducing the risk of power outages. To improve alarm processing efficiency, the power system has established a multi-channel, hierarchical information notification mechanism. These channels include, but are not limited to: standardized alarm texts generated by dispatch automation systems, operation logs generated by equipment intelligent terminals, text records submitted by operators, SMS notification platforms, and text content transcribed from telephone voice records.

[0003] Despite the increasing variety of alarm information sources, their text formats exhibit a high degree of heterogeneity and unstructured nature. For example, structured alarm data output by the dispatch system primarily uses standard templates, facilitating parsing and response; while SMS alarms are often compressed, colloquial short sentences, omitting subject-verb or parameter information; and speech-to-text may suffer from recognition errors, unclear sentence breaks, and ambiguous subject-verb agreement. These differences lead to significant variations in the expression, length, structure, and vocabulary of the same alarm event across different channels. For instance, "abnormal primary current of the main transformer" might be expressed as "high current of the main transformer" in an SMS message, while in speech-to-text it might be transcribed as "the main transformer tripped."

[0004] Furthermore, the operational scenario involves asynchrony between channels: alarm information may arrive first via voice channels and then be fed back by the system platform; some information may be repeatedly published on multiple channels, or it may be missing from some channels due to channel failure. All these characteristics constitute a complex scenario in the power alarm identification process, seriously challenging the modeling capabilities, semantic consistency capabilities, and stability of the triggering mechanism of existing systems.

[0005] (2) Existing technical solutions Most mainstream power system alarm identification technologies are based on keyword-driven shallow identification logic. The core idea is to construct a manually defined "keyword trigger library" containing keywords such as common equipment names, fault types, and operational actions. By performing string matching or regular expression matching operations on the input text, it is determined whether the corresponding alarm event is triggered. This type of method is mainly used in the following two alarm handling scenarios: 1) Structured platform alarm identification When equipment malfunctions, dispatch automation systems or monitoring platforms generate standard text alarm messages according to predetermined templates. The system can then map this text to a specific type of alarm event, such as "line trip," "switch malfunction," or "bus voltage too high," through field extraction and keyword comparison. In this process, the system relies on static field definitions and template stability for logical judgments, and essentially does not process natural language content, thus ensuring a certain level of accuracy.

[0006] 2) Matching and Recognition of SMS and Voice Alarms For SMS alerts, the system typically constructs a vocabulary lookup table and triggers judgment statements based on conditions. For example, if the SMS text contains both the keywords "bus" and "abnormal," the system matches it as an "abnormal bus voltage" event. The speech-to-text processing is similar to SMS, but due to its strong colloquial characteristics and disorganized sentence structure, the system often employs stop word filtering and word frequency matching to clean the text before performing keyword comparison. Such matching systems usually have a built-in finite state machine with several preset event templates to correspond to known fault types.

[0007] These methods face numerous challenges in practical application: 1) Literal-driven mechanisms lack semantic understanding capabilities Existing solutions rely on static keyword comparison and cannot identify variations in expression. For example, although phrases like "abnormal device current," "current tripped," and "high transformer load" describe similar events, the system cannot effectively identify them because the keywords do not completely overlap. When the input contains synonyms, near-synonyms, or abbreviations, the system typically exhibits recognition failures or misclassification.

[0008] 2) Lack of context awareness Current alarm recognition processes mostly process alarms at the sentence level, lacking cross-sentence analysis or contextual reasoning capabilities. The system struggles to determine whether alarm texts from two different channels describe the same fault scenario, and cannot deduce event content through semantic logic in the absence of explicit keywords. This is especially true in SMS and voice-to-text messages, where grammatical irregularities may lead to missing or reversed event subjects or actions, causing system processing failures.

[0009] 3) Inability to unify the modeling representation of multiple input sources Existing technical solutions typically model independently per channel, failing to support cross-channel semantic fusion. For example, dispatch system text uses field parsing, while voice text uses text cleaning and keyword extraction, lacking a unified data interface and semantic mapping mechanism. This fragmented modeling approach cannot support global judgment and response integration for alarm events, leading to frequent instances of duplicate triggering or missed responses.

[0010] 4) The process triggering mechanism relies on manual judgment. Even if some systems can complete preliminary event identification, the mapping from event to process still requires on-duty personnel to consult manuals and manually execute operating procedures. The system lacks a structured mapping system between events and process nodes, making it difficult to achieve closed-loop automated management from identification to execution, which greatly affects response time and processing consistency.

[0011] Although recent studies have attempted to incorporate Natural Language Processing (NLP) methods to improve recognition capabilities, such as TF-IDF vectors, bag-of-words models, and entity recognition, these methods are generally limited to word frequency statistics and shallow grammatical analysis, and still cannot effectively model context, sentence structure variations, and semantic logic. Therefore, in multi-source heterogeneous and highly sensitive scenarios, existing technologies cannot yet meet the closed-loop intelligent processing goal of "accurate understanding—automatic recognition—process triggering". Summary of the Invention

[0012] Based on this, it is necessary to propose an intelligent processing method, device, electronic equipment, and storage medium for relay protection to address the existing intelligent processing problems.

[0013] A smart processing method for relay protection, the method comprising: The text to be recognized is preprocessed to obtain standardized text; The standardized text is input into a pre-trained language model for semantic encoding to obtain a text encoding vector. The semantic similarity between the text encoding vector and the pre-stored event vectors in the preset standard event graph is calculated. Based on the calculation results, the pre-stored event vector with the highest similarity is selected as the target event vector; Obtain the business process diagram corresponding to the target event vector; Automated processing is performed based on the aforementioned business process diagram.

[0014] Furthermore, the step of preprocessing the text to be recognized to obtain standardized text includes: Obtain the source channel of the text to be identified; The corresponding preprocessing method is obtained based on the source channel; The text to be identified is preprocessed based on the aforementioned preprocessing method to obtain standardized text.

[0015] Furthermore, before the step of calculating the semantic similarity between the text encoding vector and the pre-stored event vectors in the preset standard event graph, the method further includes: Retrieve historical alarm records; The historical alarm records are extracted into semantic triples to obtain multiple sets of semantic triples; A preset standard event graph is constructed based on multiple sets of semantic triples.

[0016] Furthermore, after the step of calculating the semantic similarity between the text encoding vector and the pre-stored event vectors in the preset standard event graph, the method further includes: Based on the calculation results, determine whether the maximum similarity is less than the preset similarity threshold; If the maximum similarity is less than the preset similarity threshold, the text encoding vector is semantically expanded to obtain multiple expanded text encoding similar vectors. The semantic similarity between the multiple text-encoded similarity vectors and each pre-stored event vector in the preset standard event graph is calculated. Using multiple text encoding similar vectors as the text encoding, the target step and the steps following the target step are repeatedly executed; wherein, the target step is to determine whether the maximum similarity is less than a preset similarity threshold based on the calculation result; or, The maximum similarity between the text encoding similar vector and each pre-stored event vector in the preset standard event graph is greater than or equal to a preset similarity threshold.

[0017] Furthermore, after the step of automating the processing based on the business process diagram, the method further includes: Record trajectory information for automated processing; The trajectory information is converted into log information in a preset format and then saved.

[0018] Furthermore, after the step of converting the trajectory information into log information in a preset format and saving it, the method further includes: Determine whether the result of the automated processing achieves the expected effect; If the expected results are not achieved, the parameters in the pre-trained language model are adjusted or the standard event graph is revised based on the log information.

[0019] Furthermore, before the step of obtaining the business process diagram corresponding to the target event vector, the method further includes: Obtain multiple process nodes corresponding to each of the pre-stored event vectors; Design the node actions of each process node based on the pre-stored event vector; A mapping relationship is constructed based on each node and its corresponding pre-stored time vector.

[0020] A smart processing device for relay protection, the device comprising: The preprocessing module is used to preprocess the text to be recognized in order to obtain standardized text; The encoding module is used to input the standardized text into a pre-trained language model for semantic encoding to obtain a text encoding vector; The calculation module is used to calculate the semantic similarity between the text encoding vector and the pre-stored event vector in the preset standard event graph; The selection module is used to select the pre-stored event vector with the highest similarity as the target event vector based on the calculation results; The acquisition module is used to acquire the business process diagram corresponding to the target event vector; The processing module is used to perform automated processing based on the business process diagram.

[0021] An electronic device includes a memory and a processor, the memory storing a computer program that, when executed by the processor, causes the processor to perform the following steps: The text to be recognized is preprocessed to obtain standardized text; The standardized text is input into a pre-trained language model for semantic encoding to obtain a text encoding vector. The semantic similarity between the text encoding vector and the pre-stored event vectors in the preset standard event graph is calculated. Based on the calculation results, the pre-stored event vector with the highest similarity is selected as the target event vector; Obtain the business process diagram corresponding to the target event vector; Automated processing is performed based on the aforementioned business process diagram.

[0022] A computer-readable storage medium storing a computer program, which, when executed by a processor, causes the processor to perform the following steps: The text to be recognized is preprocessed to obtain standardized text; The standardized text is input into a pre-trained language model for semantic encoding to obtain a text encoding vector. The semantic similarity between the text encoding vector and the pre-stored event vectors in the preset standard event graph is calculated. Based on the calculation results, the pre-stored event vector with the highest similarity is selected as the target event vector; Obtain the business process diagram corresponding to the target event vector; Automated processing is performed based on the aforementioned business process diagram.

[0023] The beneficial effects of this invention are as follows: By generating accurate semantic encoding vectors through a pre-trained language model and selecting the event vectors with the highest similarity, the response strategy can be optimized and automated processing can be implemented, enabling rapid response to complex situations. This significantly improves the real-time performance, security, and stability of the relay protection system, reduces the burden of manual review, and enhances the overall decision support capability. Attached Figure Description

[0024] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0025] in: Figure 1 This is an application environment diagram of an intelligent processing method for relay protection in one embodiment; Figure 2 This is a flowchart of an intelligent processing method for relay protection in one embodiment; Figure 3 This is a structural block diagram of an intelligent processing device for relay protection in one embodiment; Figure 4 This is a structural block diagram of an electronic device in one embodiment. Detailed Implementation

[0026] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0027] Figure 1 This is a diagram illustrating an intelligent processing application environment for relay protection in one embodiment. (Refer to...) Figure 1 This intelligent processing method for relay protection is applied to an intelligent processing system for relay protection. The intelligent processing system for relay protection includes a terminal 110 and a server 120. The terminal 110 and server 120 are connected via a network. The terminal 110 can be a desktop terminal or a mobile terminal; the mobile terminal can be at least one of a mobile phone, tablet, or laptop. The server 120 can be a standalone server or a server cluster consisting of multiple servers. The terminal 110 is used to acquire standardized text, and the server 120 is used for automated processing.

[0028] like Figure 2As shown, in one embodiment, an intelligent processing method for relay protection is provided. This method can be applied to both terminals and servers; this embodiment illustrates its application to a terminal. The intelligent processing method for relay protection specifically includes the following steps: S1: Preprocess the text to be recognized to obtain standardized text; S2: Input the standardized text into a pre-trained language model for semantic encoding to obtain a text encoding vector; S3: Calculate the semantic similarity between the text encoding vector and the pre-stored event vector in the preset standard event graph; S4: Select the pre-stored event vector with the highest similarity as the target event vector based on the calculation results; S5: Obtain the business process diagram corresponding to the target event vector; S6: Perform automated processing based on the business process diagram.

[0029] As described in step S1 above, the text to be recognized is preprocessed to obtain standardized text. A series of preprocessing steps are performed on the text to be recognized to ensure the effectiveness and accuracy of subsequent processing. These preprocessing steps may specifically include text denoising (such as removing redundant symbols and correcting typos), terminology standardization (unifying expressions such as "trip" and "power outage" into standard action terms), named entity recognition (extracting entity words such as device name, location name, and alarm type), and format restoration. The final standardized text is characterized by clear structure, readable semantics, and consistent style. The text to be recognized can be alarm records generated by the dispatch platform, log data uploaded by devices, SMS notifications received by on-duty personnel, and text records of voice calls transcribed using Automatic Speech Recognition (ASR).

[0030] As described in step S2 above, the standardized text is input into a pre-trained language model for semantic encoding to obtain text encoding vectors. The pre-processed standardized text is then input into a pre-trained language model, such as BERT (Bidirectional Encoder Representations from Transformers) or GPT. This application preferably uses the BERT model, which possesses deep bidirectional modeling capabilities and can capture contextual dependencies between words, making it suitable for understanding irregular and semantically complex alarm descriptions. Specifically, the BERT model has been trained on a large amount of corpus and has powerful language understanding and semantic expression capabilities. By inputting the standardized text into the BERT model, the model performs deep semantic analysis of the text based on its internal neural network structure, generating high-dimensional vectors representing the text (i.e., text encoding vectors). These text encoding vectors not only contain lexical information but also imply contextual relationships and semantic information between words, effectively capturing the latent semantics of the text. Importantly, these encoding vectors can effectively reflect the similarities and differences between different components in the text, making the vectors of related texts relatively close, while the vectors of unrelated texts are relatively far apart. In addition, to further enhance BERT's adaptability in power alarm contexts, the system uses scheduling system corpus to fine-tune it, enabling it to accurately process industry-specific expressions such as device names, alarm terms, action words, and status descriptions, thereby improving semantic distinguishability.

[0031] Assuming the preprocessed text is a word sequence Then, by inputting the sequence into the BERT model, we obtain the contextual semantic embedding of each word: in For words The context representation vector, For the embedding dimension, 768 or 1024 are typically chosen. In the semantic layer, each vector not only encodes the semantic information of the word itself, but also incorporates the lexical relationships and syntactic structure before and after it, forming a semantically complete word-level representation.

[0032] The system further extracts the vector corresponding to the [CLS] bit tag. As a collective semantic representation of the entire text: This vector can be used for similarity matching with standard events in the semantic space, and it retains multi-dimensional semantic information such as event type, device object, and anomaly description. To enhance the recognizability of aggregated semantics, the system normalizes the vector after encoding and records the change in its semantic entropy value to assess whether there are redundant expressions or ambiguous information in the text, thereby assisting in the subsequent adjustment of matching confidence.

[0033] As described in step S3 above, the semantic similarity between the text-encoded vector and the pre-stored event vectors in the preset standard event map is calculated. The preset standard event map is used to represent all types of events and their structural characteristics that may occur in relay protection scenarios. This map is based on the event triplet structure and adopts the (subject-action-state / location) representation method to decompose the core information of alarm events into structurally expressible constituent units. This triplet structure is uniformly defined by the expert group to ensure that there is a one-to-one semantic relationship between equipment expression, action type and state category, thereby reducing semantic overlap and mismatch. The obtained text-encoded vector is compared with all pre-stored event vectors in the standard event map to calculate their semantic similarity. Commonly used similarity calculation methods include cosine similarity, Euclidean distance or Manhattan distance, etc. Cosine similarity is preferred because it can reflect the angle between vectors rather than the specific absolute value, which is suitable for measurement under high-dimensional sparse data.

[0034] in, The semantic vector is the encoded input text. For the first event in the standard event graph The semantic representation of each event is given by the numerator being the dot product of two vectors and the denominator being the Euclidean norm of each vector. express and The semantic similarity is calculated using this formula. This formula ensures stable comparisons even with different input lengths and encoding scales.

[0035] As described in step S4 above, the pre-stored event vector with the highest similarity is selected as the target event vector based on the calculation results. After completing the similarity calculation, it is determined which pre-stored event vector is semantically closest to the text to be identified. By sorting the similarities obtained above, the event vector with the highest similarity value is selected as the final target event vector. In one embodiment, a specific similarity threshold needs to be considered. If the highest similarity is lower than a certain set threshold, it may mean that there are no related events in the standard event graph of the text to be identified, which may require additional processing or the generation of new events.

[0036] As described in steps S5-S6 above, the business process diagram corresponding to the target event vector is obtained, and automated processing is performed based on the business process diagram. After successfully identifying the target event vector, the flowchart associated with the event vector is obtained from the preset business process diagram. The identified standard events will serve as inputs for process-driven operations, mapped to the system's predefined business process diagram, realizing coordinated control from semantic recognition to operational response. An event-process node mapping table is pre-built, establishing a one-to-many mapping relationship between each standard event and several process action nodes to represent the response path between events and automated processing steps. For example: in Indicates the event The triggered The system comprises several automated process nodes, with typical nodes including: alarm classification and judgment, voice prompts for on-duty personnel, remote control command generation for relay protection equipment, graphical monitoring highlighting, operation log archiving, and self-test command issuance. Upon event identification, the system immediately retrieves the corresponding process link from a mapping table. This link is stored as a directed graph, with nodes having associated weights, execution priorities, and precondition identifiers, supporting complex process patterns such as concurrent control, sequential flow, and conditional branching. The scheduling module loads this link into the process scheduling engine and triggers nodes sequentially according to the defined control logic. This makes the entire event handling process smoother and more intelligent, significantly improving operational efficiency and ultimately achieving intelligent relay protection.

[0037] In one embodiment, step S1, which preprocesses the text to be recognized to obtain standardized text, includes: S101: Obtain the source channel of the text to be identified; S102: Obtain the corresponding preprocessing method based on the source channel; S103: The text to be identified is preprocessed based on the preprocessing method to obtain standardized text.

[0038] As described in steps S101-S103 above, the source channel of the text to be identified is obtained. The source channels mainly include alarm records generated by the dispatch platform, log data uploaded by the device, SMS notifications received by on-duty personnel, and text records transcribed from voice calls using Automatic Speech Recognition (ASR). Since these input sources differ in language style, information completeness, text length, and grammatical structure, unified standardization processing must be completed before semantic modeling. Identify the text source channel and establish targeted preprocessing rules in advance. For structured platform text, field labels need to be parsed and non-natural language data removed; for SMS input, omitted structures need to be completed and colloquial words eliminated; for ASR-transcribed text, common speech recognition errors need to be corrected and punctuation and semantic boundaries restored. After standardization (the first preprocessing step), the second preprocessing step can be performed, which involves uniformly performing the following operations on the standardized text: text denoising (such as removing redundant symbols and correcting typos), terminology standardization (unifying expressions such as "trip" and "power outage" into standard action terms), named entity recognition (extracting entity words such as device name, location name, and alarm type), and format restoration, ultimately generating a standardized text sequence with clear structure, readable semantics, and consistent style.

[0039] In one embodiment, before step S3 of calculating the semantic similarity between the text encoding vector and pre-stored event vectors in a preset standard event graph, the method further includes: S201: Retrieve historical alarm records; S202: Extract semantic triples from the historical alarm records to obtain multiple sets of semantic triples; S203: Construct a preset standard event graph based on multiple sets of semantic triples.

[0040] As described in steps S201-S203 above, to achieve semantic normalization recognition of alarm texts, a structured standard event semantic graph needs to be constructed beforehand to represent all types of events and their structural characteristics that may occur in relay protection scenarios. Specifically, historical alarm records are obtained from the system database or relevant information storage. These historical alarm records are collections of previously occurring event data, typically containing information about problems, anomalies, or operational failures. By obtaining these records, the system can conduct in-depth analysis of past events, thereby providing a basis for subsequent event identification, pattern discovery, and preprocessing. Historical alarm records not only help understand the types and frequencies of past events but also provide guidance for predicting and handling strategies for future events. Furthermore, the analysis of historical data can help identify potential risks and patterns, thereby optimizing the system's monitoring and response mechanisms. In practice, it is necessary to ensure the completeness and accuracy of historical alarm records to avoid biases in subsequent analysis due to missing or incorrect data. At the same time, considering privacy and security, relevant data protection regulations must be followed when obtaining historical alarm records to ensure that user data is not processed in violation of regulations. The obtained historical alarm records will lay the foundation for subsequent semantic extraction and graph construction. After successfully acquiring historical alarm records, semantic triples are extracted from these records. The aim is to transform each alarm message into a structured knowledge representation, typically consisting of three parts: subject, predicate, and object. A (subject-action-state / location) representation is used to decompose the core information of the alarm event into structured components. The extraction process can utilize Natural Language Processing (NLP) techniques, such as Named Entity Recognition (NER) and relation extraction, to identify key entities in the alarm records and the relationships between them. Alternatively, a unified definition by an expert group can ensure a one-to-one semantic correspondence between device representations, action types, and state categories, thereby reducing semantic overlap and mismatches. After extracting multiple sets of semantic triples from historical alarm records, a pre-defined standard event graph is constructed based on these triples. The event graph is a knowledge network where nodes represent different event categories or entities, and edges represent the relationships and interactions between them. Semantic triples provide an intuitive representation of the connections between different events. The standard event graph is constructed through manual analysis of historical alarm records, summarizing equipment fault templates, and integrating the experience of power experts. It covers hundreds of typical event types, and each triple has a unique identifier. To maintain the consistency and up-to-dateness of the graph, the system has a graph supplementation mechanism that supports the periodic import of new event structures and automatically verifies their repetition and semantic distance from existing triples to avoid graph redundancy.

[0041] To ensure the comparability of standard events, they also need to be transformed into semantic vectors. Therefore, the system defines a unified template sentence, such as "Device X experiences state Z at location Y," and the triples are filled into the template before being input into the BERT model for encoding. in, This represents the i-th triplet. This represents the structured representation of the i-th triplet. This indicates that the structured i-th triple is input into the BERT model for encoding. This represents the i-th triple after encoding. Each standard event vector retains its structural semantic features and is mapped to the same embedding space as the input text. During this process, the system diversifies the template sentence structure, introducing multiple expressions (e.g., "Y of device X appears with Z") to enhance the generalization ability of the event graph. Simultaneously, the encoded standard event vectors are stored in a vector database, along with their corresponding event IDs and structural labels, facilitating efficient subsequent retrieval and matching operations.

[0042] Through the construction and encoding of this map, the system not only achieves semantic normalization of power alarm events, but also establishes a semantic reference standard, providing an accurate, comparable, and structured semantic foundation for subsequent matching and identification mechanisms.

[0043] In one embodiment, after step S3 of calculating the semantic similarity between the text encoding vector and pre-stored event vectors in a preset standard event graph, the method further includes: S401: Determine whether the maximum similarity is less than the preset similarity threshold based on the calculation results; S402: If the maximum similarity is less than the preset similarity threshold, then the text encoding vector is semantically expanded to obtain multiple expanded text encoding similar vectors; S403: Calculate the semantic similarity between the multiple text-encoded similarity vectors and each pre-stored event vector in the preset standard event graph; S404: Using multiple text encoding similarity vectors as the text encoding, repeatedly execute the target step and the steps following the target step; wherein, the target step is to determine whether the maximum similarity is less than a preset similarity threshold based on the calculation result; or, The maximum similarity between the text encoding similar vector and each pre-stored event vector in the preset standard event graph is greater than or equal to a preset similarity threshold.

[0044] As described in steps S401-S403 above, the calculated similarity results are evaluated to determine whether the maximum similarity value is lower than a preset similarity threshold. The similarity threshold is set based on the needs of the actual application scenario and empirical data, representing the minimum requirement for determining the association between text and standard events. If the maximum similarity value is less than this threshold, it indicates a weak semantic connection between the text to be identified and the standard event, meaning the event described by the text to be identified is not within the scope of the preset standard event map. Known similar words or relevant information appearing in the same context are used to generate multiple extended text-encoded similarity vectors. The semantic neighborhood expansion method may also utilize techniques such as nearest neighbor generation, synonym generation, and contextual similarity in the word embedding space to create text feature vectors that better reflect the topic or scenario. This introduces a dynamic semantic neighborhood expansion strategy. When the initial matching does not meet the confidence requirement, the set of semantic points within a certain radius around the input vector can be aggregated to form a "soft boundary matching," mitigating expression shifts caused by word order reversal, expression omission, and passive verb conversion. This mechanism can effectively identify synonymous, ambiguous, and vague expressions, such as recognizing phrases like "power outage," "tripped circuit breaker," and "device unresponsive" as the same event. Furthermore, because it uses vector distance for comparison and does not rely on exact keyword matching, it has stronger expressive adaptability, especially suitable for the expressive uncertainties brought about by colloquial or non-standard input sources (such as SMS messages and speech-to-text transcription). During operation, the system also records the similarity distribution, Top-K candidates, and the final selected event number for each recognition task, storing these as recognition logs in the background database. This provides a training foundation for subsequent model optimization, anomaly analysis, and knowledge graph expansion. After successfully performing semantic neighborhood expansion and obtaining multiple text-encoded similarity vectors, the semantic association between each expanded text-encoded similarity vector and the standard event vector is re-evaluated.

[0045] As described in step S404 above, the multiple text encoding similarity vectors calculated in the previous step are used as new text encodings, and the previously defined target steps are executed again to re-determine and calculate similarity. This allows for continuous expansion and iteration to improve the accuracy and effectiveness of recognition, even when the initial text encoding is insufficient. If, after similarity calculation, the maximum similarity of a certain similar vector is higher than or equal to a preset threshold, this event can be confirmed and used as the basis for processing. However, if there are still multiple similar vectors whose maximum similarity is lower than the threshold, the system will again perform the semantic neighborhood expansion process, iterating continuously until a suitable matching result is obtained. In this way, the system implements an adaptive learning and updating mechanism, which can continuously adjust the recognition strategy when the data changes dynamically, improving the overall processing accuracy and efficiency.

[0046] In one embodiment, after step S6 of automating the process based on the business process diagram, the method further includes: S701: Records trajectory information for automated processing; S702: Convert the trajectory information into log information in a preset format and save it.

[0047] As described in steps S701-S702 above, after the process is completed, the system will record the complete processing trajectory information of this alarm, including the original input content, preprocessing standardization results, semantic vector number, matching score, matching event identifier, trigger process node sequence, execution status and operation results of each node, etc. The log is stored in a structured format and retains the original semantic input, supports indexing and backtracking, and provides a basis for subsequent anomaly identification analysis and graph maintenance.

[0048] In one embodiment, after step S702 of converting the trajectory information into log information in a preset format and saving it, the method further includes: S7031: Determine whether the result of the automated processing achieves the expected effect; S7032: If the expected results are not achieved, adjust the parameters in the pre-trained language model or revise the standard event graph based on the log information.

[0049] As described in steps S7031-S7032 above, to improve the system's adaptability and self-evolution level, this invention designs a feedback update mechanism based on abnormal samples. When manual review or process execution does not conform to expectations, the system will automatically mark the processing record as a potential misidentification match and guide the map revision or semantic model parameter adjustment. Low-confidence matching samples will enter the offline training pool for periodically updating the BERT embedding model, or dynamically correct the matching threshold according to statistical strategies, improving the model's responsiveness to changes in alarm semantics. Through the closed-loop process of "recording—identifying—feedback—updating," the system ensures the accuracy of identification and the reliability of control response when facing new events or expression variations.

[0050] In one embodiment, before step S5 of obtaining the business process diagram corresponding to the target event vector, the method further includes: S411: Obtain multiple process nodes corresponding to each of the pre-stored event vectors; S412: Design the node actions of each process node based on the pre-stored event vector; S413: Construct a mapping relationship based on each node and its corresponding pre-stored time vector.

[0051] As described in steps S411-S413 above, multiple corresponding process nodes are extracted and obtained for each pre-stored event vector. A process node is the basic building block in a business process diagram. An event-process node mapping table is pre-built, establishing a one-to-many mapping relationship between each standard event and several process action nodes to represent the response path between the event and the automated processing steps. For example: in Indicates the event The triggered The system comprises several automated process nodes, with typical nodes including: alarm classification and judgment, voice prompts for on-duty personnel, remote control command generation for relay protection equipment, graphical monitoring highlighting, operation log archiving, and self-test command issuance. Upon event identification, the system immediately invokes a mapping table to locate its corresponding process link. This link is stored as a directed graph, with nodes having associated weights, execution priorities, and precondition identifiers, supporting complex process modes such as concurrent control, sequential flow, and conditional branching. The scheduling module loads this link into the process scheduling engine and triggers nodes sequentially according to the defined control logic. This makes the entire event handling process smoother and more intelligent, significantly improving operational efficiency and ultimately achieving intelligent relay protection. After designing the corresponding node actions for each process node, a mapping relationship is established between each node and its corresponding pre-stored event vector, tightly connecting events with the actual processing flow. This ensures that whenever a specific event occurs, the corresponding business process node and its actions can be quickly located, directly entering the corresponding processing path, providing a foundation for subsequent automated processing.

[0052] Reference Figure 3 The present invention also provides an intelligent processing device for relay protection, the device comprising: The preprocessing module 902 is used to preprocess the text to be recognized in order to obtain standardized text; Encoding module 904 is used to input the standardized text into a pre-trained language model for semantic encoding to obtain a text encoding vector; The calculation module 906 is used to calculate the semantic similarity between the text encoding vector and the pre-stored event vector in the preset standard event graph; Module 908 is used to select the pre-stored event vector with the highest similarity as the target event vector based on the calculation results. The acquisition module 910 is used to acquire the business process diagram corresponding to the target event vector; Processing module 912 is used for automated processing based on the business process diagram.

[0053] In one embodiment, the preprocessing module 902 includes: The source channel acquisition submodule is used to acquire the source channel of the text to be identified; The preprocessing method acquisition submodule is used to acquire the corresponding preprocessing method based on the source channel; The preprocessing submodule is used to preprocess the text to be identified based on the preprocessing method to obtain standardized text.

[0054] In one embodiment, the intelligent processing device for relay protection further includes: The historical alarm record acquisition module is used to acquire historical alarm records; The semantic triple extraction module is used to extract semantic triples from the historical alarm records to obtain multiple sets of semantic triples. A preset standard event graph construction module is used to construct a preset standard event graph based on multiple sets of semantic triples.

[0055] In one embodiment, the intelligent processing device for relay protection further includes: The similarity judgment module is used to determine whether the maximum similarity is less than a preset similarity threshold based on the calculation results; The semantic neighborhood expansion module is used to expand the semantic neighborhood of the text encoding vector if the maximum similarity is less than a preset similarity threshold, so as to obtain multiple expanded text encoding similar vectors. The semantic similarity calculation module is used to calculate the semantic similarity between the multiple text-encoded similarity vectors and each pre-stored event vector in the preset standard event graph; An iterative module is used to repeatedly execute the target step and the steps following the target step, using multiple text encoding similarity vectors as the text encoding; wherein, the target step is to determine whether the maximum similarity is less than a preset similarity threshold based on the calculation result; or, The maximum similarity between the text encoding similar vector and each pre-stored event vector in the preset standard event graph is greater than or equal to a preset similarity threshold.

[0056] In one embodiment, the intelligent processing device for relay protection further includes: The trajectory information recording module is used to record trajectory information processed automatically. The log information saving module is used to convert the trajectory information into log information in a preset format and save it.

[0057] In one embodiment, the intelligent processing device for relay protection further includes: The expected effect judgment module is used to determine whether the result after the automated processing achieves the expected effect; The adjustment module is used to adjust the parameters in the pre-trained language model or revise the standard event graph based on the log information if the expected results are not achieved.

[0058] In one embodiment, the intelligent processing device for relay protection further includes: The process node acquisition module is used to acquire multiple process nodes corresponding to each of the pre-stored event vectors; The node action design module is used to design node actions for each process node based on the pre-stored event vector. The mapping relationship construction module is used to construct a mapping relationship based on each node and its corresponding pre-stored time vector.

[0059] Figure 4 An internal structural diagram of an electronic device in one embodiment is shown. This electronic device can specifically be a terminal or a server, and more specifically, a computer device. Figure 4 As shown, the electronic device includes a processor, a memory, and a network interface connected via a system bus. The memory includes a non-volatile storage medium and internal memory. The non-volatile storage medium stores an operating system and may also store a computer program. When executed by the processor, this computer program enables the processor to implement an intelligent processing method for relay protection. The internal memory may also store a computer program, which, when executed by the processor, enables the processor to implement an intelligent processing method for relay protection. Those skilled in the art will understand that... Figure 4 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the electronic device to which the present application is applied. The specific electronic device may include more or fewer components than shown in the figure, or combine certain components, or have different component arrangements.

[0060] In one embodiment, an electronic device is provided, including a memory and a processor, the memory storing a computer program that, when executed by the processor, causes the processor to perform the following steps: The text to be recognized is preprocessed to obtain standardized text; The standardized text is input into a pre-trained language model for semantic encoding to obtain a text encoding vector. The semantic similarity between the text encoding vector and the pre-stored event vectors in the preset standard event graph is calculated. Based on the calculation results, the pre-stored event vector with the highest similarity is selected as the target event vector; Obtain the business process diagram corresponding to the target event vector; Automated processing is performed based on the aforementioned business process diagram.

[0061] By generating accurate semantic encoding vectors through pre-trained language models and selecting the event vectors with the highest similarity, response strategies can be optimized and automated processing can be implemented to quickly respond to complex situations. This significantly improves the real-time performance, security, and stability of relay protection systems, reduces the burden of manual review, and enhances overall decision support capabilities.

[0062] In one embodiment, a computer-readable storage medium is provided storing a computer program that, when executed by a processor, causes the processor to perform the following steps: The text to be recognized is preprocessed to obtain standardized text; The standardized text is input into a pre-trained language model for semantic encoding to obtain a text encoding vector. The semantic similarity between the text encoding vector and the pre-stored event vectors in the preset standard event graph is calculated. Based on the calculation results, the pre-stored event vector with the highest similarity is selected as the target event vector; Obtain the business process diagram corresponding to the target event vector; Automated processing is performed based on the aforementioned business process diagram.

[0063] By generating accurate semantic encoding vectors through pre-trained language models and selecting the event vectors with the highest similarity, response strategies can be optimized and automated processing can be implemented to quickly respond to complex situations. This significantly improves the real-time performance, security, and stability of relay protection systems, reduces the burden of manual review, and enhances overall decision support capabilities.

[0064] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments described above. Any references to memory, storage, databases, or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), RAMbus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and RAMbus dynamic RAM (RDRAM), etc.

[0065] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0066] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of this patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this patent application should be determined by the appended claims.

Claims

1. A method of intelligent processing of a relay protection, characterized in that, The method comprises: preprocessing the text to be identified to obtain standardized text; inputting the standardized text into a pre-trained language model for semantic coding to obtain a text coding vector; calculating the semantic similarity between the text coding vector and the pre-stored event vector in the preset standard event graph; selecting the pre-stored event vector with the highest similarity as the target event vector according to the calculation result; obtaining the business process graph corresponding to the target event vector; automatically processing based on the business process graph.

2. The intelligent processing method of claim 1, wherein, The step of preprocessing the text to be identified to obtain standardized text comprises: obtaining the source channel of the text to be identified; obtaining the corresponding preprocessing mode based on the source channel; preprocessing the text to be identified based on the preprocessing mode to obtain standardized text.

3. The intelligent processing method of claim 1, wherein, Before the step of calculating the semantic similarity between the text coding vector and the pre-stored event vector in the preset standard event graph, the method further comprises: obtaining historical alarm records; extracting semantic triples from the historical alarm records to obtain multiple groups of semantic triples; constructing a preset standard event graph based on multiple groups of semantic triples.

4. The intelligent processing method of claim 1, wherein, After the step of calculating the semantic similarity between the text coding vector and the pre-stored event vector in the preset standard event graph, the method further comprises: determining whether the maximum similarity is less than a preset similarity threshold according to the calculation result; if the maximum similarity is less than the preset similarity threshold, performing semantic neighborhood expansion on the text coding vector to obtain multiple extended text coding similar vectors; calculating the semantic similarity between the multiple text coding similar vectors and each pre-stored event vector in the preset standard event graph; repeating the target step and the steps after the target step, wherein the target step is to determine whether the maximum similarity is less than the preset similarity threshold according to the calculation result; or, satisfying that the maximum similarity between the text coding similar vector and each pre-stored event vector in the preset standard event graph is greater than or equal to the preset similarity threshold.

5. The intelligent processing method of claim 1, wherein, After the step of automatically processing based on the business process graph, the method further comprises: recording the trajectory information of the automatic processing; converting the trajectory information into log information in a preset format and saving it.

6. The intelligent processing method of claim 5, wherein, After the step of converting the trajectory information into log information in a preset format and saving it, the method further comprises: determining whether the result of the automatic processing achieves the expected effect; if the expected effect is not achieved, adjusting the parameters in the pre-trained language model or revising the standard event graph according to the log information.

7. The intelligent processing method of claim 1, wherein, Before the step of obtaining the business process graph corresponding to the target event vector, the method further comprises: obtaining multiple process nodes corresponding to each pre-stored event vector; designing node actions for each process node based on the pre-stored event vector; constructing a mapping relationship between each node and the corresponding pre-stored event vector.

8. An intelligent processing device for protection relaying, characterized by The device comprises: a preprocessing module for preprocessing the text to be identified to obtain standardized text; The encoding module is configured to input the standardized text into a pre-trained language model for semantic encoding to obtain a text encoding vector; The computing module is configured to calculate semantic similarity between the text encoding vector and a pre-stored event vector in a pre-set standard event graph; The selecting module is configured to select, according to a calculation result, a pre-stored event vector with the highest similarity as a target event vector; The obtaining module is configured to obtain a business process graph corresponding to the target event vector; The processing module is configured to perform automatic processing based on the business process graph.

9. A computer-readable storage medium, characterized in that, The device comprises a memory and a processor, and the memory stores a computer program which, when executed by the processor, causes the processor to perform the steps of the intelligent processing method for relay protection according to any one of claims 1 to 7.

10. An electronic device, comprising: The device comprises a memory and a processor, and the memory stores a computer program which, when executed by the processor, causes the processor to perform the steps of the intelligent processing method for relay protection according to any one of claims 1 to 7.

Images