Substation equipment intelligent monitoring method and system based on guangming big model

Abstract

The present application relates to the technical field of intelligent monitoring of power equipment, in particular to a substation equipment intelligent monitoring method and system based on a Guangming big model, comprising: obtaining unstructured original alarm signal text of a substation, removing redundant information through text cleaning and standardization segmentation processing, and extracting core signal name fragments; inputting the fragments into a Guangming big model semantic understanding engine to generate a deep semantic feature vector, synchronously reading a standard feature set of a standardized knowledge base, fusing the deep semantic feature vector with vocabulary and structural features to obtain a comprehensive feature representation, and then comparing and calculating the comprehensive similarity value with the standard feature set one by one to finally complete automatic classification decision, output corresponding equipment type and alarm level identification, and support subsequent control and display operation. The present application optimizes the processing and classification process of substation alarm signals through directional text processing and big model semantic feature fusion.

Description

Technical Field

[0001] This invention relates to the field of intelligent monitoring technology for power equipment, and in particular to an intelligent monitoring method and system for substation equipment based on the Guangming large model. Background Technology

[0002] Currently, substation equipment monitoring primarily employs rule-based matching and keyword retrieval to process alarm signal text generated by monitoring systems. Existing technologies directly extract and compare features from the complete original alarm text, relying on preset fixed rules to filter signal content and completing signal matching and classification through single words or structural features. They fail to perform targeted redundant information removal for unstructured alarm text. Under this approach, punctuation characters, separators, and common prefixes in the original alarm text continuously interfere with the signal recognition process, making it impossible to accurately extract pure core signal name fragments. Furthermore, relying solely on shallow word and structural features for matching lacks a large-scale semantic understanding mechanism, failing to convert signal names into deep semantic feature vectors, and failing to achieve comprehensive similarity calculation after multi-feature fusion.

[0003] Redundant information in unstructured alarm signal text from substations can cause biases in signal feature recognition. Fixed rules and single-feature comparison methods cannot capture the deep semantic connotations of signal names, and the accuracy of signal matching and classification is limited by the unstructured nature of the text. This invention addresses the problems of not being able to remove redundant information and extract pure core signal name fragments from unstructured original alarm signal text, as well as the inability to generate deep semantic feature vectors through large models and to complete multi-feature fusion and comprehensive similarity comparison calculations. It constructs a corresponding signal processing and classification decision-making process. Summary of the Invention

[0004] The purpose of this invention is to overcome the shortcomings of existing technologies and propose an intelligent monitoring method and system for substation equipment based on the Guangming large model.

[0005] To achieve the above objectives, the present invention adopts the following technical solution: a substation equipment intelligent monitoring method based on the Guangming large model, comprising: Obtain the original alarm signal text generated by the substation monitoring system. The original alarm signal text contains unstructured equipment identification and status description information. The original alarm signal text is cleaned and standardized for segmentation. Redundant punctuation characters, delimiters and common prefixes are removed, and the core signal name fragments are extracted. The core signal name fragment is input into the preset Guangming Big Model semantic understanding engine, which converts the core signal name fragment into a corresponding deep semantic feature vector. The deep semantic feature vector is used to characterize the semantic connotation of the signal name. The standard feature set of all known signal entries is read synchronously from a pre-built standardized knowledge base. The standard feature set includes the standard semantic vector, standard lexical features and standard structural features of the known signal entries. The deep semantic feature vector of the current signal is fused with its own lexical features and structural features to generate a comprehensive feature representation of the current signal. The comprehensive feature representation of the current signal is compared one by one with the standard feature set of all known signal entries in the standardized knowledge base to obtain a series of comprehensive similarity values. Automatic classification decisions are made based on the aforementioned series of comprehensive similarity values.

[0006] As a further aspect of the present invention, text cleaning and standardized segmentation processing is performed on the original alarm signal text to remove redundant punctuation characters, delimiters, and common prefixes, and to extract the core signal name fragment, including: The original alarm signal text is scanned at the character level to identify and delete all punctuation characters, including colons, semicolons, hyphens, and underscores. In the text after removing punctuation characters, based on the preset general prefix word list for substation equipment, the general prefix words located at the beginning of the text are matched and removed; The remaining text after removing common prefixes is segmented according to logical delimiters commonly used in signal naming, which consist of space characters or specific keywords. From the segmented text, select text units containing the actual device component names and status / action descriptions, and recombine the selected text units in their original order to form the core signal name segment; The core signal name fragments are uniformly converted to lowercase character format to eliminate the differences in lexical features caused by inconsistent capitalization.

[0007] As a further aspect of the present invention, the step of synchronously reading the standard feature set of all known signal entries from a pre-constructed standardized knowledge base includes: Access the central storage node of the standardized knowledge base, which contains a set of known signal entries that have been classified and labeled. Each known signal entry is associated with a device type identifier and an alarm level identifier. From each known signal entry, read the standard semantic vector generated after the signal name has been transformed by the Guangming Big Model semantic understanding engine; From each known signal entry, the inherent lexical pattern of its signal name text is extracted as the standard lexical feature, which includes a set of abbreviations for specific equipment models and standard operating terms; From each known signal entry, the structural pattern of its signal name text is parsed as the standard structural feature, which includes a description of the number of words in the signal name and the word order. The standard semantic vector, the standard lexical features, and the standard structural features are logically bound to the device type identifier and alarm level identifier corresponding to the known signal entry, and together they constitute the standard feature set of the known signal entry.

[0008] As a further aspect of the present invention, the step of fusing the deep semantic feature vector of the current signal with its own lexical features and structural features to generate a comprehensive feature representation of the current signal includes: The core signal name segment of the current signal is segmented and part-of-speech tagging is performed to identify the set of words belonging to power industry terminology, and the set of words is encoded as the lexical features of the current signal. The total number of words in the core signal name segment is counted, and the position order of each word in the segment is recorded. The sequence of the total number of words and the position order is encoded as the structural feature of the current signal. The deep semantic feature vector, lexical features, and structural features of the current signal are normalized to ensure that their values ​​are within the same dimension range. A weighted concatenation method is used to connect the normalized deep semantic feature vector, the lexical features, and the structural features into a higher-dimensional fusion feature vector; The fused feature vector is subjected to dimensionality reduction and compression to generate a comprehensive feature representation of the current signal with fixed dimensions.

[0009] As a further aspect of the present invention, the comprehensive feature representation of the current signal is compared one by one with the standard feature set of all known signal entries in the standardized knowledge base to obtain a series of comprehensive similarity values, including: For each known signal entry in the standardized knowledge base, extract the standard fusion feature vector contained in its standard feature set; Calculate the cosine similarity between the comprehensive feature representation of the current signal and the standard fused feature vector to obtain a preliminary semantic similarity value; Calculate the Jaccard similarity coefficient between the lexical features of the current signal and the standard lexical features of the currently known signal entry to obtain a lexical similarity value; Compare the structural features of the current signal with the standard structural features of the currently known signal entries, calculate the degree of matching between the structural features of the current signal and the standard structural features of the currently known signal entries in terms of total number of words and positional order, and obtain a structural similarity value; The calculated semantic similarity, lexical similarity, and structural similarity values ​​are weighted and summed to obtain a comprehensive similarity value that represents the overall similarity between the current signal and the currently known signal entries.

[0010] As a further aspect of the present invention, automatic classification decision-making based on the aforementioned series of comprehensive similarity values ​​includes: Traverse the series of comprehensive similarity values, find the comprehensive similarity value with the largest value, and record the known signal entry corresponding to the largest comprehensive similarity value as the target reference entry; Read the currently effective comprehensive similarity confidence threshold from the system configuration parameters; Compare the largest overall similarity value with the overall similarity confidence threshold; When the maximum comprehensive similarity value is greater than or equal to the comprehensive similarity confidence threshold, the associated device type identifier and alarm level identifier are read from the attributes of the target reference entry. The read device type identifier and alarm level identifier are used as the classification result and assigned to the currently processed original alarm signal text to complete the automatic classification.

[0011] As a further aspect of the present invention, it also includes triggering a self-correction process of the standardized knowledge base when automated classification fails: When the maximum comprehensive similarity value is less than the comprehensive similarity confidence threshold, the automated classification decision is deemed to have failed. Retrieve all known signal entries from the standardized knowledge base whose comprehensive feature representation similarity to the current signal exceeds an initial high threshold, and form a high-confidence similar signal set. The initial high threshold is set as the mean of historical similarity values ​​in the knowledge base plus three times the standard deviation. Statistically analyze the frequency of occurrence of each device type identifier and alarm level identifier in the set of high-confidence similar signals; Select the most frequently occurring device type identifier and alarm level identifier as the recommended classification label for the current signal, and generate a correction suggestion entry containing the original signal text, the recommended classification label, and evidence of similar signals; The proposed corrections and the set of high-confidence similar signals are submitted to the automated verification queue, triggering an automated verification process based on preset logic rules. Once the suggested correction entry passes the automated verification process, the current signal, its finally confirmed device type identifier, alarm level identifier, and the deep semantic feature vector obtained from the Guangming Big Model Semantic Understanding Engine are stored as a new known signal entry in the standardized knowledge base, thus completing the incremental update of the knowledge base.

[0012] As a further aspect of the present invention, it also includes the step of iteratively reducing the similarity threshold to expand the coverage of the knowledge base correction: After completing one round of signal correction and knowledge base update under the initial high threshold condition, the updated state of the knowledge base is used as the knowledge source for the next round of correction. Read the preset threshold descent step size from the system configuration parameters, and generate an extended matching threshold that is lower than the initial high threshold based on the threshold descent step size; Using the extended matching threshold as a new condition, the remaining unclassified signals that failed to match due to insufficient similarity in the previous round are retrieved again in the updated standardized knowledge base; For each of the remaining unclassified signals, under the new extended matching threshold conditions, its comprehensive similarity with known signal entries in the knowledge base is recalculated, and a new round of retrieval, statistics, recommendation, automated verification process and knowledge base update process is executed. Repeat the cycle of knowledge source update, matching threshold decrease by step size, retrieval and correction update until the matching threshold drops to the preset minimum threshold, or the number of remaining unmatched signals is lower than the preset number and they are different from each other, then terminate the automatic correction iteration process. The step of using the expanded matching threshold as a new condition to re-retrieve the remaining unclassified signals that failed to match in the previous round due to insufficient similarity from the updated standardized knowledge base specifically includes: At the beginning of each iteration, load the complete data of the standardized knowledge base that was updated after the previous iteration; Load the list of remaining unclassified signals recorded after the end of the previous iteration, for which no similar known signal entries could be found; Using the matching threshold that is effective in the current round, recalculate the overall similarity value between each of the remaining unclassified signals in the list and all known signal entries in the standardized knowledge base; For each of the remaining unclassified signals, all known signal entries whose comprehensive similarity value exceeds the matching threshold of the current round are selected to form the candidate similar signal set for the current round; If the candidate similar signal set of a certain remaining unclassified signal is not empty, the remaining unclassified signal is marked as a processable signal in the current round and enters the frequency statistics and label recommendation process; if its candidate similar signal set is empty, it is kept in the list of remaining unclassified signals and enters the next iteration.

[0013] As a further aspect of the present invention, the construction steps of the Guangming large-scale model semantic understanding engine include: Historical alarm texts and standardized signal name texts from the field of power equipment operation monitoring were collected to form the original training corpus. The original training corpus is preprocessed, including text cleaning, word segmentation, and noise filtering. The preprocessed corpus is divided into a training set, a validation set, and a test set for model training and performance evaluation. Define the basic network architecture of the large model, which includes a multi-layer Transformer encoder and a self-attention mechanism module; The training set is used to perform unsupervised masked language model pre-training on the basic network architecture of the large model, enabling the model to learn the general semantic representation of text in the power field; Using the labeled dataset corresponding to the power equipment signal classification task, supervised fine-tuning training is performed on the pre-trained model to optimize the model's ability to encode the deep semantics of signal names. The hyperparameters during model training are adjusted using the validation set, and the final semantic understanding accuracy of the model is evaluated using the test set. The trained and evaluated model parameters are solidified and deployed as the Guangming Big Model semantic understanding engine, which can be called online, to convert the input core signal name fragments into corresponding deep semantic feature vectors.

[0014] As a further aspect of the present invention, the present invention also includes a substation equipment intelligent monitoring system based on the Guangming large model. The system includes a memory, a processor, and a computer program stored in the memory and running on the processor. When the processor executes the computer program, it implements the steps of the substation equipment intelligent monitoring method based on the Guangming large model as described above.

[0015] Compared with the prior art, the advantages and positive effects of the present invention are as follows: The original alarm signal text generated by the substation monitoring system undergoes text cleaning and standardized segmentation to remove redundant punctuation, delimiters, and common prefixes. Core signal name fragments are extracted, eliminating irrelevant and interfering content from unstructured text. This allows the extraction of signal features to focus on the core content, reducing the interference of redundant information on subsequent feature processing and making the signal name feature source purer. It also avoids feature shift problems caused by invalid content in the complete text. The core signal name fragments are then input into the Guangming Big Data Model semantic understanding engine to be converted into deep semantic feature vectors. This enables the mining of the semantic connotations behind the signal names, breaking through the limitations of traditional shallow lexical and structural feature representations. It fully presents the semantic relationships and essential meanings of the signal names, extending the representation dimension of signal features from the surface form to the deep semantic level.

[0016] By fusing deep semantic feature vectors with the signal's own lexical and structural features to generate a comprehensive feature representation, signal feature information from different dimensions can be integrated to form a more comprehensive signal feature representation, no longer limited to the presentation of information from a single feature dimension. By comparing the comprehensive feature representation of the current signal with the standard feature set of known signal entries in a standardized knowledge base one by one to calculate the comprehensive similarity value, simultaneous comparison of multi-dimensional features can be achieved, making the basis for signal matching more comprehensive. Automatic classification decisions are completed based on the comprehensive similarity value, ensuring that the judgment logic for signal classification aligns with the complete feature attributes of the signal, overcoming the one-sidedness of traditional single-feature comparison, and optimizing the fit and rationality of classification decisions. Attached Figure Description

[0017] Figure 1 The flowchart shows the intelligent monitoring method for substation equipment based on the Guangming large model described in this invention. Figure 2 A flowchart for reading a set of standard features from a standardized knowledge base; Figure 3 A line graph showing the relationship between similarity matching threshold and signal classification success rate; Figure 4 This is a diagram illustrating the threshold adjustment and iterative correction analysis. Figure 5 This is a diagram illustrating the iterative process of signal processing. Detailed Implementation

[0018] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0019] In the description of this invention, it should be understood that the terms "length," "width," "upper," "lower," "front," "rear," "left," "right," "vertical," "horizontal," "top," "bottom," "inner," and "outer," etc., indicating orientation or positional relationships, are based on the orientation or positional relationships shown in the accompanying drawings and are only for the convenience of describing the invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation, and therefore should not be construed as a limitation of the invention. Furthermore, in the description of this invention, "a plurality of" means two or more, unless otherwise explicitly specified.

[0020] See Figure 1 This invention provides an intelligent monitoring method for substation equipment based on the Guangming Big Data Model. The method includes: acquiring the original alarm signal text generated by the substation monitoring system, which contains unstructured equipment identification and status description information; performing text cleaning and standardized segmentation on the original alarm signal text to remove redundant punctuation characters, separators, and common prefixes, thereby extracting the core signal name fragment; inputting the core signal name fragment into a pre-built Guangming Big Data Model semantic understanding engine, which converts the core signal name fragment into a corresponding deep semantic feature vector, which represents the semantic connotation of the signal name; synchronously reading the standard feature set of all known signal entries from a pre-constructed standardized knowledge base, which includes the standard semantic vector, standard lexical features, and standard structural features of the known signal entries; fusing the deep semantic feature vector of the current signal with its own lexical and structural features to generate a comprehensive feature representation of the current signal; comparing the comprehensive feature representation of the current signal with the standard feature set of all known signal entries in the standardized knowledge base one by one to obtain a series of comprehensive similarity values; and making automatic classification decisions based on this series of comprehensive similarity values.

[0021] In one embodiment of the present invention, during text cleaning and standardized segmentation, the original alarm signal text is scanned at the character level to identify and delete all punctuation characters, including colons, semicolons, hyphens, and underscores. In the text after deleting punctuation, common prefixes located at the beginning of the text are matched and removed based on a preset list of common prefixes for substation equipment. The remaining text after removing common prefixes is segmented according to common logical delimiters used in signal naming; these delimiters consist of spaces or specific keywords. From the segmented text fragments, text units containing actual equipment component names and status / action descriptions are selected, and these selected text units are recombined in their original order to form core signal name fragments. Finally, the core signal name fragments are uniformly converted to lowercase character format to eliminate lexical feature differences caused by inconsistent capitalization.

[0022] In practical implementation, the original alarm signal text undergoes preprocessing. This preprocessing involves extracting core signal name fragments from the unstructured equipment identification and status description information. The preprocessing process performs a character-level scan of the original alarm signal text, identifying and deleting all punctuation characters, including colons, semicolons, hyphens, and underscores. After removing punctuation, the system performs a matching operation based on a preset general prefix word list for substation equipment. This list contains a series of common prefix words that do not carry specific equipment information. The system matches and removes successfully matched general prefix words located at the beginning of the text.

[0023] In practice, the remaining text after removing common prefixes is further segmented based on common logical delimiters used in signal naming. These logical delimiters consist of spaces or specific keywords. From the text fragments generated by the segmentation operation, the system selects text units containing actual device component names and status / action descriptions according to predefined rules. These selected text units are then recombined in their original order of appearance in the original alarm signal text to form core signal name fragments. In some embodiments, the specific keywords for the logical delimiters are words such as "fault," "action," and "abnormality." The segmentation operation can be formally defined as segmenting the text at the locations where these specific keywords or spaces are encountered. It is understood that the number of text fragments generated after segmentation is related to the frequency of occurrence of the logical delimiters. Finally, the core signal name fragments are uniformly converted to lowercase character format. This operation aims to eliminate lexical differences caused by inconsistent capitalization in English, ensuring the stability of subsequent processing.

[0024] In some embodiments, the preset general prefix glossary for substation equipment can be configured and updated according to the naming conventions of different substation monitoring systems. Optionally, in the step of deleting punctuation characters, the range of characters to be deleted can be expanded, such as parentheses, forward slashes, etc. Optionally, before segmenting according to logical delimiters, additional whitespace normalization processing can be performed on the text, merging multiple consecutive whitespace characters into a single whitespace character. It can be understood that extracting core signal name fragments is the foundation for subsequent semantic understanding and feature matching. After cleaning and standardization segmentation, the core signal name fragments are more regular, which is beneficial to improving the conversion accuracy of the Guangming big data model semantic understanding engine and the effectiveness of feature comparison. The entire processing flow is automated, and structured core signal name fragments can be extracted from the original alarm signal text without manual intervention.

[0025] In one embodiment of the present invention, the standard feature set of all known signal entries is synchronously read from a pre-built standardized knowledge base, see [reference]. Figure 2This process involves accessing the central storage node of a standardized knowledge base containing a set of known signal entries that have been categorized and labeled. Each known signal entry is associated with a device type identifier and an alarm level identifier. From each known signal entry, a standard semantic vector generated by transforming its signal name using the Guangming Big Data Model semantic understanding engine is retrieved. From each known signal entry, the inherent lexical pattern of its signal name text is extracted as a standard lexical feature, which includes a set of abbreviations for specific device models and standard operating terms. From each known signal entry, the structural pattern of its signal name text is parsed as a standard structural feature, which includes a description of the number of words in the signal name and the word order. The standard semantic vector, standard lexical feature, and standard structural feature are logically bound to the device type identifier and alarm level identifier corresponding to the known signal entry to jointly constitute the standard feature set of that known signal entry. Historical alarm texts and standardized signal name texts from the field of power equipment operation monitoring are collected to form the original training corpus. The original training corpus is preprocessed, including text cleaning, word segmentation, and noise filtering. The preprocessed corpus is divided into training, validation, and test sets for model training and performance evaluation. A basic network architecture for the large-scale model is defined, comprising a multi-layer Transformer encoder and a self-attention mechanism module. Unsupervised masked language model pre-training is performed on the basic network architecture of the large-scale model using the training set, enabling the model to learn general semantic representations of text in the power industry. Supervised fine-tuning training is then performed on the pre-trained model using the labeled dataset corresponding to the power equipment signal classification task, optimizing the model's ability to encode deep semantics of signal names. Hyperparameters are adjusted during model training using the validation set, and the final semantic understanding accuracy of the model is evaluated using the test set. The trained and evaluated model parameters are then solidified and deployed as an online-accessible semantic understanding engine for converting input core signal name fragments into corresponding deep semantic feature vectors.

[0026] In practical implementation, the standard feature set of all known signal entries is synchronously read from a pre-built standardized knowledge base. The construction and reading of the standard feature set is the foundation for classification and comparison. This process is achieved by accessing the central storage node of the standardized knowledge base, which is a structured database containing a set of known signal entries that have been classified and labeled. Each known signal entry not only contains the signal name text but also logically associates it with a device type identifier and an alarm level identifier. From each known signal entry, the system reads the standard semantic vector generated after the signal name text has been offline converted by the Guangming Big Model Semantic Understanding Engine. The standard semantic vector is a high-dimensional floating-point vector used to encode the deep semantic information of the signal name. Simultaneously, from the same known signal entry, the inherent lexical pattern of its signal name text is extracted as standard lexical features. Standard lexical features are a set of words that include specific device model abbreviations and standard operating terms identified from the signal name. In some embodiments, standard lexical features may also include a hash-encoded proper noun index.

[0027] In practical implementation, in addition to standard semantic vectors and standard lexical features, the system also parses the structural patterns of signal name texts from known signal entries as standard structural features. Standard structural features describe the composition of signal names, including a digital description of the total number of words in the signal name and the order of each word. Optionally, standard structural features can be represented using a sequence, where each position records the part-of-speech tag or category code of the corresponding word. It can be understood that logically binding the standard semantic vectors, standard lexical features, and standard structural features with the device type identifier and alarm level identifier corresponding to the known signal entry constitutes the standard feature set of that known signal entry. When signal classification is required, the system synchronously reads the standard feature sets of all known signal entries in the standardized knowledge base, providing a complete comparison benchmark for subsequent comprehensive similarity calculations. In practical implementation, the construction process begins with the collection of domain corpus, collecting historical alarm texts from the power equipment operation monitoring field and standardized signal name texts compiled by experts to form the original training corpus. Preprocessing operations are performed on the original training corpus, including text cleaning, word segmentation, and rule-based noise filtering. The preprocessed corpus is divided into training, validation, and test sets according to a predetermined ratio. The training set is used for model parameter learning, the validation set for hyperparameter tuning, and the test set for final performance evaluation. A basic network architecture for the Guangming model is defined, consisting of a multi-layer Transformer encoder and a self-attention mechanism module. The model parameters are randomly initialized. The basic network architecture is then pre-trained using the training set in an unsupervised masked language model. The goal of this pre-training is to enable the model to predict randomly masked words, thereby allowing it to learn general semantic representations of text in the power industry.

[0028] In some embodiments, after pre-training, the pre-trained model is subjected to supervised fine-tuning training using the labeled dataset corresponding to the power equipment signal classification task. Fine-tuning optimizes the model's ability to encode the deep semantics of signal names through a classification layer, ensuring that the semantic vectors generated by the model are closer together in the vector space for signals of the same class and farther apart for signals of different classes. The loss function for fine-tuning can be cross-entropy loss, expressed as: in: The average loss value of the batch. Represents the number of samples in the batch. The total number of signal categories. It is an indicator function, when the sample The true category is The value is 1 if it is true, and 0 otherwise. Representative model predicts samples Category The probability of obtaining the model's semantic understanding is determined. Hyperparameters such as learning rate and batch size are adjusted during model training using a validation set, and the semantic understanding accuracy of the trained Guangming Big Model semantic understanding engine is evaluated using an independent test set. Optionally, evaluation metrics may include accuracy and recall for vector similarity tasks. Finally, the trained and evaluated model parameters are solidified and deployed as an online service, namely the Guangming Big Model semantic understanding engine, which can receive input core signal name fragments and output corresponding deep semantic feature vectors.

[0029] In one embodiment of the present invention, a comprehensive feature representation of the current signal is generated. The core signal name segment of the current signal is segmented and part-of-speech tagging is performed to identify the set of words belonging to power industry terminology, and this set of words is encoded as the lexical features of the current signal. The total number of words in the core signal name segment is counted, and the positional order of each word in the segment is recorded. The sequence of the total number of words and their positional order is encoded as the structural features of the current signal. The deep semantic feature vector, lexical features, and structural features of the current signal are normalized to ensure their values ​​are within the same dimension. A weighted concatenation method is used to connect the normalized deep semantic feature vector, lexical features, and structural features into a higher-dimensional fusion feature vector. The fusion feature vector is then subjected to dimensionality reduction and compression to generate a comprehensive feature representation of the current signal with a fixed dimension. A comprehensive similarity value is calculated. For each known signal entry in the standardized knowledge base, the standard fusion feature vector contained in its standard feature set is extracted. The cosine similarity between the comprehensive feature representation of the current signal and the standard fusion feature vector is calculated to obtain a preliminary semantic similarity value. Calculate the Jaccard similarity coefficient between the lexical features of the current signal and the standard lexical features of the known signal entries to obtain a lexical similarity score. Compare the structural features of the current signal with the standard structural features of the known signal entries, calculating their matching degree in terms of total number of words and positional order to obtain a structural similarity score. Finally, perform a weighted sum of the calculated semantic similarity score, lexical similarity score, and structural similarity score to obtain a comprehensive similarity score representing the overall similarity between the current signal and the known signal entries.

[0030] In the specific implementation, the comprehensive feature representation of the current signal is generated. This process begins with the core signal name segment of the current signal. The core signal name segment undergoes word segmentation and part-of-speech tagging. Using a pre-trained professional dictionary and parser, the set of words belonging to the power industry terminology is identified. This set of power industry terminology is encoded as the lexical features of the current signal. Lexical features can be represented as the sum of the one-hot encoded vectors of these words or the average of the word embedding vectors. Simultaneously, the total number of words in the core signal name segment is counted, and the positional index of each word within the core signal name segment is recorded. The total number of words and the sequence of positional indices are jointly encoded as the structural features of the current signal. Structural features can be represented as a combination of an integer and an array of positional indices. In the specific implementation, the deep semantic feature vector, lexical features, and structural features of the current signal are normalized. Normalization uses min-max scaling or Z-score standardization methods to ensure that the values ​​of the deep semantic feature vector, lexical features, and structural features are within the same dimensional range, eliminating the bias caused by different original dimensions of the features in subsequent fusion. A weighted concatenation method is used to connect the normalized deep semantic feature vector, normalized lexical features, and normalized structural features into a higher-dimensional fusion feature vector. Before concatenation, each feature vector is multiplied by a preset weight coefficient. The resulting fusion feature vector is then subjected to dimensionality reduction and compression. Principal component analysis or autoencoder techniques can be used to generate a comprehensive feature representation of the current signal with a fixed dimension lower than the original dimension of the fusion feature vector.

[0031] In practice, the comprehensive feature representation of the current signal is compared one by one with the standard feature set of all known signal entries in the standardized knowledge base. For each known signal entry in the standardized knowledge base, the standard fusion feature vector contained in the known signal entry is extracted from the standard feature set. The standard fusion feature vector is obtained by performing the same normalization, weighted concatenation, and dimensionality reduction processing on the standard semantic vector, standard lexical features, and standard structural features during the knowledge base construction phase. The cosine similarity between the comprehensive feature representation of the current signal and the standard fusion feature vector is calculated. The cosine similarity calculation is based on the dot product and magnitude of the two vectors, resulting in a preliminary semantic similarity value between -1 and 1. The Jaccard similarity coefficient between the lexical features of the current signal and the standard lexical features of the known signal entries is calculated. The Jaccard similarity coefficient is defined as the ratio of the number of elements in the intersection to the number of elements in the union of the two lexical feature sets, resulting in a lexical similarity value between 0 and 1. The structural features of the current signal are compared with the standard structural features of known signal entries. The degree of matching between the current signal's structural features and the standard structural features of known signal entries is calculated in terms of total number of words and positional order. The matching degree calculation considers the absolute difference in the total number of words and the consistency of words in the same position, resulting in a structural similarity score. In a specific calculation formula, the structural similarity score... It can be represented as: in: This represents the calculated structural similarity value. The total number of words representing the current signal. The total number of standard terms representing currently known signal entries. It is a constant that adjusts the decay rate. This represents the word length of the shorter of the two signal names. Represents the current signal of the first Word identifiers at each location, The first in the representative standard signal Word identifiers at each location, It is a comparison function, when equal The value is 1 when the time is right, and 0 otherwise. and It is a weighting coefficient, and The calculated semantic similarity, lexical similarity, and structural similarity values ​​are weighted and summed. Each similarity value in the weighted summation is multiplied by a preset weight coefficient, resulting in a comprehensive similarity value that represents the overall similarity between the current signal and the known signal entries. The comprehensive similarity value ranges from 0 to 1. In some embodiments, the weight coefficients used in feature fusion and similarity calculation can be determined on a validation set using grid search or optimization algorithms. Examples of preset values ​​are shown in Table 1.

[0032] Table 1: Weighting Coefficients for Feature and Similarity Calculation Optionally, when calculating the structural similarity score, the weight of the term representing the difference in the total number of words can be adjusted. It can be set to 0.4, the weight of the position order matching item. It can be set to 0.6. Optionally, the Jaccard similarity coefficient calculation for lexical features is defined as 0 when the lexical feature set is empty. It can be understood that the comprehensive similarity value integrates information from three dimensions: semantics, vocabulary, and structure. It can also be understood that weighted summation is a linear combination method; in specific deployments, the weight coefficients need to be adjusted based on the actual classification performance.

[0033] See Figure 3 This is a line graph showing the relationship between the similarity matching threshold and the signal classification success rate, reflecting the key principles of iterative optimization of the knowledge base in a substation equipment intelligent monitoring system. The two are significantly negatively correlated: the lower the threshold, the higher the classification success rate; conversely, the higher the threshold, the lower the success rate. When the threshold is 0.950, the success rate is only about 72%, matching only extremely similar known signals, leaving many marginal signals unclassified. When the threshold drops to 0.750, the success rate approaches 97%, covering more similar / marginal signals and significantly improving classification coverage. The success rate increases rapidly in the early stages of threshold reduction, but slows down later, indicating that there is still room for improvement in classification at low thresholds, but the marginal effect diminishes. In the high-threshold stage, only highly similar signals are matched, resulting in low classification coverage and a low success rate. In the iterative threshold reduction stage, medium-to-low similarity signals are gradually matched, and the knowledge base is supplemented through automated verification, achieving incremental updates.

[0034] In one embodiment of the present invention, automatic classification decision-making is performed based on a series of comprehensive similarity values. The series of comprehensive similarity values ​​is traversed, and the comprehensive similarity value with the largest value is found. The known signal entry corresponding to the largest comprehensive similarity value is recorded as the target reference entry. The currently effective comprehensive similarity confidence threshold is read from the system configuration parameters. The comprehensive similarity value with the comprehensive similarity confidence threshold is compared. When the comprehensive similarity value with the largest value is greater than or equal to the comprehensive similarity confidence threshold, the associated device type identifier and alarm level identifier are read from the attributes of the target reference entry. The read device type identifier and alarm level identifier are assigned as the classification result to the currently processed original alarm signal text, completing the automatic classification. The equipment type identifier and alarm level identifier obtained after the automated classification is completed, i.e., the final classification result, will be used as a definite output and immediately applied to the subsequent technical response process. The final classification result is transmitted in real time to the substation's automation control subsystem. The automation control subsystem has a pre-stored control strategy table that logically maps different equipment types and alarm levels. Upon receiving the identifier, it immediately queries the table and generates the corresponding equipment control command sequence. This command sequence is sent to the corresponding physical actuators in the substation, such as relay protection devices, circuit breakers, disconnect switches, etc., to drive them to perform predetermined opening, closing, blocking, or parameter adjustment operations, thereby realizing physical feedback or isolation of abnormal equipment states. The final classification results are also synchronously transmitted to the monitoring system's human-machine interface (HMI). Based on the alarm level identifier in the results, the HMI automatically triggers a specific alarm rendering strategy bound to that level. This strategy includes, but is not limited to, highlighting faulty equipment elements on the main wiring diagram or equipment topology diagram using a highlighted or flashing format; displaying the faulty equipment at the top of the real-time alarm list with a preset color and font distinct from other levels; and automatically linking and displaying the standard emergency response plan, historical fault records, or operation ticket interface for that type of equipment. The semantic information contained in the original alarm signal text is transformed into executable control commands and structured monitoring interface interaction information through semantic understanding and classification methods. This enables indirect or direct control of substation physical equipment and significantly improves the efficiency of alarm recognition and handling by maintenance personnel. When automated classification fails, a self-correction process of the standardized knowledge base is triggered. When the maximum comprehensive similarity value is less than the comprehensive similarity confidence threshold, the automated classification decision is deemed to have failed. Retrieve all known signal entries from the standardized knowledge base whose comprehensive feature representation similarity to the current signal exceeds an initial high threshold, forming a high-confidence similar signal set. The initial high threshold is set as the mean of historical similarity values ​​in the knowledge base plus three standard deviations. Statistically analyze the frequency of each device type identifier and alarm level identifier within the high-confidence similar signal set.The most frequently occurring device type identifier and alarm level identifier are selected as recommended classification labels for the current signal, and a correction suggestion entry containing the original signal text, recommended classification labels, and evidence of similar signals is generated. The correction suggestion entry and a set of high-confidence similar signals are submitted to an automated verification queue, triggering an automated verification process based on preset logical rules. Once the correction suggestion entry passes the automated verification process, the current signal, its finally confirmed device type identifier, alarm level identifier, and the deep semantic feature vector obtained from the Guangming Big Data Model semantic understanding engine are stored as a new known signal entry in the standardized knowledge base, completing the incremental update of the knowledge base.

[0035] In practical implementation, automatic classification decisions are made based on a series of comprehensive similarity values. The system iterates through the current signal and compares it with all known signal entries in the standardized knowledge base to generate a series of comprehensive similarity values. It then identifies the comprehensive similarity value with the highest value and records the known signal entry corresponding to this highest value as the target reference entry. The system reads the currently effective comprehensive similarity confidence threshold from the system configuration parameters and compares the largest comprehensive similarity value with the comprehensive similarity confidence threshold. When the largest comprehensive similarity value is greater than or equal to the comprehensive similarity confidence threshold, the system reads the device type identifier and alarm level identifier associated with the target reference entry from its attributes. This read device type identifier and alarm level identifier are then assigned as the classification result to the currently processed original alarm signal text, thus completing the automated classification. In some embodiments, the comprehensive similarity confidence threshold can be set to 0.85. The system configuration parameters allow administrators to dynamically adjust the comprehensive similarity confidence threshold based on actual operating results. Optionally, when multiple comprehensive similarity values ​​are tied for the highest, the secondary similarity between the standard feature sets corresponding to these tied entries and the current signal features can be further compared, or one can be randomly selected as the target reference entry. When the largest comprehensive similarity value is less than the comprehensive similarity confidence threshold, the system determines that the automated classification decision has failed and triggers a self-correction process in the standardized knowledge base. The self-correction process retrieves all known signal entries from the standardized knowledge base whose comprehensive feature representation similarity to the current signal exceeds an initial high threshold, forming a high-confidence similar signal set. Initial high threshold The setting depends on the historical data of the knowledge base. Specifically, it can be set as the mean of the comprehensive similarity scores of all historical successfully matched records in the knowledge base plus three standard deviations. The formula is expressed as: in: The arithmetic mean of the historical comprehensive similarity scores. The standard deviation represents the historical comprehensive similarity values. The frequency of each device type identifier and alarm level identifier in the high-confidence similar signal set is statistically analyzed. The device type identifier and alarm level identifier with the highest frequency are selected as the recommended classification labels for the current signal, and a correction suggestion entry is generated. The correction suggestion entry includes the original signal text, the recommended classification label, and a list of high-confidence similar signal sets as supporting evidence. The correction suggestion entry and the high-confidence similar signal set are submitted to the automated verification queue, triggering an automated verification process based on preset logical rules. After the correction suggestion entry passes the automated verification process, the system stores the current signal, its finally confirmed device type identifier, alarm level identifier, and the deep semantic feature vector obtained from the Guangming Big Model semantic understanding engine as a new known signal entry in the standardized knowledge base, completing the incremental update of the knowledge base. It can be understood that the automated verification process is a crucial step in ensuring the quality of knowledge base updates. It can be understood that the standardized knowledge base achieves continuous expansion and optimization during operation through a self-correction process. See Table 2, which shows the frequency statistics of a high-confidence similar signal set.

[0036] Table 2: Frequency Statistics of High-Confidence Similar Signal Sets In some embodiments, if the set of high-confidence similar signals is empty, the automated verification process will initiate a sub-process for difference analysis and temporary category label generation according to preset new category creation rules, or mark it as a special case awaiting expert review and store it in an independent queue, without interrupting the automated operation of the main process. Optionally, when making a decision, the automated verification process can simultaneously retrieve and analyze the statistical frequency and difference between the recommended classification label (i.e., the most frequent label) and the second most frequent label, using this as an input parameter for preset verification rules to evaluate the confidence level and decision boundary of the classification recommendation.

[0037] See Figure 4 This is a threshold adjustment and iterative correction analysis diagram, fully presenting the changes in threshold adjustment and system performance during the iterative correction of the knowledge base of the substation intelligent monitoring system. Initially, the high threshold resulted in extremely high accuracy but extremely low coverage, with many edge signals failing to match. The first adjustment increased coverage by 10 percentage points, while accuracy decreased slightly by 3 percentage points. The second adjustment saw a continuous increase in coverage, while accuracy decreased linearly. The third adjustment resulted in the convergence of coverage and accuracy, reaching a performance balance. The lowest threshold reached its peak coverage, while accuracy dropped to an acceptable lower limit. By gradually lowering the similarity threshold, the system's matching range expanded, and coverage significantly improved, but this introduced more edge signals, leading to a slight decrease in accuracy. In the "third adjustment" stage, both coverage and accuracy reached 88%, representing the node with the highest business value. This resolved the signal backlog problem under the initial high threshold while avoiding the risk of excessively low accuracy under the lowest threshold.

[0038] In one embodiment of the present invention, the method further includes a step of iteratively reducing the similarity threshold to expand the coverage of the knowledge base correction. After completing one round of signal correction and knowledge base update under the initial high threshold condition, the updated state of the knowledge base is used as the knowledge source for the new round of correction. A preset threshold descent step size is read from the system configuration parameters, and an extended matching threshold lower than the initial high threshold is generated based on the threshold descent step size. Using the extended matching threshold as the new condition, the remaining unclassified signals that failed to match in the previous round due to insufficient similarity are re-retrieved in the updated standardized knowledge base. For each remaining unclassified signal, its comprehensive similarity with known signal entries in the knowledge base is recalculated under the new extended matching threshold condition, and a new round of retrieval, statistics, recommendation, automated verification process and knowledge base update process is executed. The cycle of knowledge source update, step-wise reduction of the matching threshold, retrieval and correction update is repeated until the matching threshold drops to a preset minimum threshold, or the number of remaining unmatched signals is lower than a preset number and they are different from each other, at which point the automatic correction iteration process is terminated. Using an expanded matching threshold as a new condition, the updated standardized knowledge base is used to re-retrieve remaining unclassified signals that failed to match in the previous round due to insufficient similarity. Specifically, at the beginning of each iteration, the complete data of the standardized knowledge base updated after the previous iteration is loaded. A list of all remaining unclassified signals that failed to find any similar known signal entries, recorded after the end of the previous iteration, is loaded. Using the matching threshold effective in the current round, the comprehensive similarity value between each remaining unclassified signal in the list and all known signal entries in the standardized knowledge base is recalculated. For each remaining unclassified signal, all known signal entries with a comprehensive similarity value exceeding the matching threshold of the current round are selected, forming the candidate similar signal set for the current round. If the candidate similar signal set of a remaining unclassified signal is not empty, the remaining unclassified signal is marked as a processable signal for the current round and enters the frequency statistics and label recommendation process; if its candidate similar signal set is empty, it is retained in the list of remaining unclassified signals and enters the next iteration.

[0039] In practice, the similarity threshold is lowered iteratively to expand the coverage of the knowledge base correction. After completing one round of signal correction and knowledge base update under the initial high threshold condition, the system uses the updated state of the knowledge base as the knowledge source for the next round of correction. The preset threshold descent step size is read from the system configuration parameters, and an expanded matching threshold lower than the initial high threshold is generated based on this step size. The formula for calculating the expanded matching threshold is: in: This represents the extended matching threshold used in the new iteration. This represents the matching threshold used in the previous iteration; in the first iteration, it is the initial high threshold. , This represents the preset fixed threshold descent step size. Using the calculated extended matching threshold as the new condition, the system re-retrieves the remaining unclassified signals that failed to match in the previous iteration due to insufficient similarity from the updated standardized knowledge base. For each remaining unclassified signal, the system recalculates the comprehensive similarity between the remaining unclassified signal and all known signal entries in the standardized knowledge base under the new extended matching threshold condition, and executes a new round of retrieval, statistics, recommendation, automated verification, and knowledge base update processes. The cycle of knowledge source update, matching threshold descent by step size, retrieval, and correction update is repeated until the matching threshold drops to the system's preset minimum threshold, or the number of remaining unmatched signals is lower than the preset number and they are different from each other. At this point, the system terminates the automatic correction iteration process.

[0040] In practice, using an expanded matching threshold as a new condition, the system re-retrieves remaining unclassified signals that failed to match in the previous round due to insufficient similarity from the updated standardized knowledge base. This involves the following steps: At the start of each iteration, the system loads the complete data of the standardized knowledge base updated after the previous iteration, ensuring the current retrieval is based on the latest knowledge base content. The system loads a list of all remaining unclassified signals that failed to find any similar known signal entries after the end of the previous iteration. Using the matching threshold effective for the current round, the system recalculates the comprehensive similarity value between each remaining unclassified signal in the list and all known signal entries in the standardized knowledge base. For each remaining unclassified signal, the system selects all known signal entries whose comprehensive similarity value exceeds the matching threshold for the current round, forming the candidate similar signal set for the current round. If the candidate similar signal set for a remaining unclassified signal is not empty, the remaining unclassified signal is marked as a processable signal for the current round and enters the frequency statistics and tag recommendation process; if the candidate similar signal set for a remaining unclassified signal is empty, the remaining unclassified signal is retained in the list of remaining unclassified signals and enters the next iteration. In some embodiments, the preset minimum threshold can be set to 0.5, and the threshold decreasing step size can be... The threshold can be set to 0.1. Optionally, the termination condition "the number of remaining unmatched signals is less than a preset number and they differ from each other" can be determined by calculating the average distance between the feature vectors of the remaining signals. It is understood that by iteratively lowering the threshold, the system can progressively process signals with low similarity to the knowledge base, expanding the coverage of automatic correction. In some embodiments, the correction suggestion entries generated and submitted in each iteration are marked with their respective iteration rounds. Optionally, the automated verification process can be configured to comprehensively consider multi-dimensional information such as the current round, the matching threshold, and the consistency of recommendation tags from historical rounds, and make a comprehensive judgment based on preset decision rules. It is understood that this iterative process achieves fully automated, adaptive, batch expansion of the knowledge base.

[0041] See Figure 5 This is a signal processing iteration analysis diagram, showcasing the signal processing data throughout the entire iterative correction process of the knowledge base in the intelligent monitoring system for substation equipment. It clearly presents the gradual convergence from "unclassified signals" to "automated verification processes." From round 1 to round 3, the number of unclassified signals decreased sharply from 500 to 160, and the number of automated verification processes decreased from 160 to 80. This demonstrates that the strategy of iteratively reducing the similarity threshold effectively incorporates a large number of marginal signals into the system's matching range, achieving an efficient transformation from "unclassified signals" to "processable signals." In each iteration, the ratio of processable signals to unclassified signals continuously increases, indicating that with incremental updates to the knowledge base, the system's ability to identify similar signals continuously improves, and the success rate of automatic processing steadily increases. The decision-making burden of the automated verification process decreases round by round; the number of signals to be processed in round 3 is only 50% of that in round 1, proving that the system achieves self-optimization through iteration, with the efficiency of automated decision-making and the completeness of the knowledge base improving simultaneously.

[0042] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention in any other way. Any person skilled in the art may make changes or modifications to the above-disclosed technical content to create equivalent embodiments that can be applied to other fields. However, any simple modifications, equivalent changes, and modifications made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the protection scope of the present invention.

Claims

1. A method for intelligent monitoring of substation equipment based on the Guangming large-scale model, characterized in that, include: Obtain the original alarm signal text generated by the substation monitoring system. The original alarm signal text contains unstructured equipment identification and status description information. The original alarm signal text is cleaned and standardized for segmentation. Redundant punctuation characters, delimiters and common prefixes are removed, and the core signal name fragments are extracted. The core signal name fragment is input into the preset Guangming Big Model semantic understanding engine, which converts the core signal name fragment into a corresponding deep semantic feature vector. The deep semantic feature vector is used to characterize the semantic connotation of the signal name. The standard feature set of all known signal entries is read synchronously from a pre-built standardized knowledge base. The standard feature set includes the standard semantic vector, standard lexical features and standard structural features of the known signal entries. The deep semantic feature vector of the current signal is fused with its own lexical features and structural features to generate a comprehensive feature representation of the current signal. The comprehensive feature representation of the current signal is compared one by one with the standard feature set of all known signal entries in the standardized knowledge base to obtain a series of comprehensive similarity values. Automatic classification decisions are made based on the series of comprehensive similarity values ​​to obtain the corresponding device type identifier and alarm level identifier, and subsequent control and display operations are performed based on the device type identifier and alarm level identifier.

2. The intelligent monitoring method for substation equipment based on the Guangming large model according to claim 1, characterized in that, The original alarm signal text is cleaned and standardized for segmentation, removing redundant punctuation, delimiters, and common prefixes to extract the core signal name fragment, including: The original alarm signal text is scanned at the character level to identify and delete all punctuation characters, including colons, semicolons, hyphens, and underscores. In the text after removing punctuation characters, based on the preset general prefix word list for substation equipment, the general prefix words located at the beginning of the text are matched and removed; The remaining text after removing common prefixes is segmented according to logical delimiters commonly used in signal naming, which consist of space characters or specific keywords. From the segmented text, select text units containing the actual device component names and status / action descriptions, and recombine the selected text units in their original order to form the core signal name segment; The core signal name fragments are uniformly converted to lowercase character format to eliminate the differences in lexical features caused by inconsistent capitalization.

3. The intelligent monitoring method for substation equipment based on the Guangming large model according to claim 2, characterized in that, The process of synchronously reading the standard feature set of all known signal entries from a pre-built standardized knowledge base includes: Access the central storage node of the standardized knowledge base, which contains a set of known signal entries that have been classified and labeled. Each known signal entry is associated with a device type identifier and an alarm level identifier. From each known signal entry, read the standard semantic vector generated after the signal name has been transformed by the Guangming Big Model semantic understanding engine; From each known signal entry, the inherent lexical pattern of its signal name text is extracted as the standard lexical feature, which includes a set of abbreviations for specific equipment models and standard operating terms; From each known signal entry, the structural pattern of its signal name text is parsed as the standard structural feature, which includes a description of the number of words in the signal name and the word order. The standard semantic vector, the standard lexical features, and the standard structural features are logically bound to the device type identifier and alarm level identifier corresponding to the known signal entry, and together they constitute the standard feature set of the known signal entry.

4. The intelligent monitoring method for substation equipment based on the Guangming large model according to claim 3, characterized in that, The step of fusing the deep semantic feature vector of the current signal with its own lexical features and structural features to generate a comprehensive feature representation of the current signal includes: The core signal name segment of the current signal is segmented and part-of-speech tagging is performed to identify the set of words belonging to power industry terminology, and the set of words is encoded as the lexical features of the current signal. The total number of words in the core signal name segment is counted, and the position order of each word in the segment is recorded. The sequence of the total number of words and the position order is encoded as the structural feature of the current signal. The deep semantic feature vector, lexical features, and structural features of the current signal are normalized to ensure that their values ​​are within the same dimension range. A weighted concatenation method is used to connect the normalized deep semantic feature vector, the lexical features, and the structural features into a higher-dimensional fusion feature vector; The fused feature vector is subjected to dimensionality reduction and compression to generate a comprehensive feature representation of the current signal with fixed dimensions.

5. The intelligent monitoring method for substation equipment based on the Guangming large model according to claim 4, characterized in that, The comprehensive feature representation of the current signal is compared one by one with the standard feature set of all known signal entries in the standardized knowledge base to obtain a series of comprehensive similarity values, including: For each known signal entry in the standardized knowledge base, extract the standard fusion feature vector contained in its standard feature set; Calculate the cosine similarity between the comprehensive feature representation of the current signal and the standard fused feature vector to obtain a preliminary semantic similarity value; Calculate the Jaccard similarity coefficient between the lexical features of the current signal and the standard lexical features of the currently known signal entry to obtain a lexical similarity value; Compare the structural features of the current signal with the standard structural features of the currently known signal entries, calculate the degree of matching between the structural features of the current signal and the standard structural features of the currently known signal entries in terms of total number of words and positional order, and obtain a structural similarity value; The calculated semantic similarity, lexical similarity, and structural similarity values ​​are weighted and summed to obtain a comprehensive similarity value that represents the overall similarity between the current signal and the currently known signal entries.

6. The intelligent monitoring method for substation equipment based on the Guangming large model according to claim 5, characterized in that, Automatic classification decisions are made based on the aforementioned series of comprehensive similarity values ​​to obtain corresponding device type identifiers and alarm level identifiers. Subsequent control and display operations are then performed based on these identifiers, including: Traverse the series of comprehensive similarity values, find the comprehensive similarity value with the largest value, and record the known signal entry corresponding to the largest comprehensive similarity value as the target reference entry; Read the currently effective comprehensive similarity confidence threshold from the system configuration parameters; Compare the largest overall similarity value with the overall similarity confidence threshold; When the maximum comprehensive similarity value is greater than or equal to the comprehensive similarity confidence threshold, the associated device type identifier and alarm level identifier are read from the attributes of the target reference entry. The read device type identifier and alarm level identifier are used as the classification result and assigned to the currently processed original alarm signal text to complete the automatic classification and obtain the final classification result; The equipment type identifier and alarm level identifier in the final classification result are transmitted to the substation's automation control subsystem. The automation control subsystem generates a corresponding equipment control command sequence based on the equipment type identifier and alarm level identifier. The equipment control command sequence is used to drive the physical actuators in the substation. The device type identifier and alarm level identifier in the final classification result are transmitted to the monitoring system human-machine interface. Based on the alarm level identifier, the monitoring system human-machine interface triggers the corresponding specific alarm rendering strategy to highlight and display the alarm information on the interface.

7. The intelligent monitoring method for substation equipment based on the Guangming large model according to claim 6, characterized in that, It also includes a self-correcting process for the standardized knowledge base when automated classification fails: When the maximum comprehensive similarity value is less than the comprehensive similarity confidence threshold, the automated classification decision is deemed to have failed. Retrieve all known signal entries from the standardized knowledge base whose comprehensive feature representation similarity to the current signal exceeds an initial high threshold, and form a high-confidence similar signal set. The initial high threshold is set as the mean of historical similarity values ​​in the knowledge base plus three times the standard deviation. Statistically analyze the frequency of occurrence of each device type identifier and alarm level identifier in the set of high-confidence similar signals; Select the most frequently occurring device type identifier and alarm level identifier as the recommended classification label for the current signal, and generate a correction suggestion entry containing the original signal text, the recommended classification label, and evidence of similar signals; The proposed corrections and the set of high-confidence similar signals are submitted to the automated verification queue, triggering an automated verification process based on preset logic rules. Once the suggested correction entry passes the automated verification process, the current signal, its finally confirmed device type identifier, alarm level identifier, and the deep semantic feature vector obtained from the Guangming Big Model Semantic Understanding Engine are stored as a new known signal entry in the standardized knowledge base, thus completing the incremental update of the knowledge base.

8. The intelligent monitoring method for substation equipment based on the Guangming large model according to claim 7, characterized in that, It also includes the step of iteratively lowering the similarity threshold to expand the coverage of the knowledge base correction: After completing one round of signal correction and knowledge base update under the initial high threshold condition, the updated state of the knowledge base is used as the knowledge source for the next round of correction. Read the preset threshold descent step size from the system configuration parameters, and generate an extended matching threshold that is lower than the initial high threshold based on the threshold descent step size; Using the extended matching threshold as a new condition, the remaining unclassified signals that failed to match due to insufficient similarity in the previous round are retrieved again in the updated standardized knowledge base; For each of the remaining unclassified signals, under the new extended matching threshold conditions, its comprehensive similarity with known signal entries in the knowledge base is recalculated, and a new round of retrieval, statistics, recommendation, automated verification process and knowledge base update process is executed. Repeat the cycle of knowledge source update, matching threshold decrease by step size, retrieval and correction update until the matching threshold drops to the preset minimum threshold, or the number of remaining unmatched signals is lower than the preset number and they are different from each other, then terminate the automatic correction iteration process. The step of using the expanded matching threshold as a new condition to re-retrieve the remaining unclassified signals that failed to match in the previous round due to insufficient similarity from the updated standardized knowledge base specifically includes: At the beginning of each iteration, load the complete data of the standardized knowledge base that was updated after the previous iteration; Load the list of remaining unclassified signals recorded after the end of the previous iteration, for which no similar known signal entries could be found; Using the matching threshold that is effective in the current round, recalculate the overall similarity value between each of the remaining unclassified signals in the list and all known signal entries in the standardized knowledge base; For each of the remaining unclassified signals, all known signal entries whose comprehensive similarity value exceeds the matching threshold of the current round are selected to form the candidate similar signal set for the current round; If the candidate similar signal set of a certain remaining unclassified signal is not empty, the remaining unclassified signal is marked as a processable signal in the current round and enters the frequency statistics and label recommendation process; if its candidate similar signal set is empty, it is kept in the list of remaining unclassified signals and enters the next iteration.

9. The intelligent monitoring method for substation equipment based on the Guangming large model according to claim 8, characterized in that, The construction steps of the Guangming Big Model semantic understanding engine include: Historical alarm texts and standardized signal name texts from the field of power equipment operation monitoring were collected to form the original training corpus. The original training corpus is preprocessed, including text cleaning, word segmentation, and noise filtering. The preprocessed corpus is divided into a training set, a validation set, and a test set for model training and performance evaluation. Define the basic network architecture of the large model, which includes a multi-layer Transformer encoder and a self-attention mechanism module; The training set is used to perform unsupervised masked language model pre-training on the basic network architecture of the large model, enabling the model to learn the general semantic representation of text in the power field; Using the labeled dataset corresponding to the power equipment signal classification task, supervised fine-tuning training is performed on the pre-trained model to optimize the model's ability to encode the deep semantics of signal names. The hyperparameters during model training are adjusted using the validation set, and the final semantic understanding accuracy of the model is evaluated using the test set. The trained and evaluated model parameters are solidified and deployed as the Guangming Big Model semantic understanding engine, which can be called online, to convert the input core signal name fragments into corresponding deep semantic feature vectors.

10. A substation equipment intelligent monitoring system based on the Guangming large model, comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the intelligent monitoring method for substation equipment based on the Guangming large model as described in any one of claims 1 to 9.

Images