A substation text auditing method, system, device and medium based on a large model

Abstract

The application relates to the technical field of large models, in particular to a power transformation text auditing method, system and device based on a large model and a medium. The application first performs professional preprocessing on power transformation text data to construct a training data set, then trains a model by using a low-rank adaptation and thought chain combination technology, and finally executes an automatic auditing process to generate a result. The application improves the accuracy and efficiency of text auditing, has strong expansibility, can flexibly cope with changing business demands, and effectively saves resources for manual review.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of large models, in particular to a power transformation text auditing method, system, device and medium based on a large model. BACKGROUND

[0002] With the continuous advancement of smart grid construction, the types and quantities of substation equipment are rapidly increasing, and the operation and maintenance management work is becoming increasingly complex. Substation operation, maintenance, and maintenance work involve a large amount of professional text data, and the accuracy and standardization of these data are directly related to the safe and stable operation of the power system.

[0003] Currently, power transformation text processing mainly relies on automated software systems to assist in completing basic functions such as text format checking, key information extraction, and data consistency verification. These systems are mainly based on traditional natural language processing techniques and perform preliminary structured analysis and processing of text, but lack the ability to understand power transformation professional terminology and complex semantics, resulting in low processing accuracy and the need for extensive manual review. Therefore, further improvements are needed. SUMMARY

[0004] To solve the problem of low processing accuracy and extensive manual review of existing power transformation text, the present application provides a power transformation text auditing method, system, device and medium based on a large model, which adopts the following technical solutions:

[0005] In a first aspect, the present application provides a power transformation text auditing method based on a large model, comprising the following steps:

[0006] Pretreating the power transformation text data to obtain a training data set;

[0007] Training a model using low-rank adaptation and thought chain combination technology based on the training data set to obtain a text auditing model;

[0008] Based on the text auditing model, performing a power transformation automation auditing process to obtain an auditing result.

[0009] By adopting the above technical solutions, due to the rapid increase in the types and quantities of substation equipment, the number of professional text data that power transformation operation and maintenance personnel need to process has increased dramatically. These data include operation procedures, operation tickets, maintenance plans, and each file must strictly follow professional standards and safety requirements. The present application first pretreats the power transformation text data to construct a training data set, then trains a model using low-rank adaptation and thought chain combination technology, and finally performs an automated auditing process to generate results. This improves the accuracy and efficiency of text auditing and has strong scalability, allowing for flexible response to changing business needs and effective saving of resources for manual review.

[0010] Optionally, the power transformation text material includes power transformation regulations and operation manuals; the power transformation text material is preprocessed to obtain a training data set, specifically including the following steps:

[0011] The power transformation regulations and operation manual files are read and standardized to obtain structured original text;

[0012] Based on the structured original text, term features are extracted and a knowledge graph is constructed to establish a term mapping relationship table to obtain term index data;

[0013] The term index data is used to perform term standardization identification and replacement processing on the text to generate normalized text;

[0014] The normalized text is applied to a word segmentation algorithm and data cleaning rule to generate a standard format training data set.

[0015] By adopting the above technical solution, the power transformation regulations and operation manuals are first systematically collected as original text to construct a professional dictionary system containing term standard definition, usage specification and scene association, then the term recognition and standardization processing is performed based on the dictionary, and finally the training data set is generated through the word segmentation and cleaning steps; the deep understanding and standardization processing of the power transformation professional text are realized, which not only improves the data quality, but also provides a reliable professional knowledge base for model training.

[0016] Optionally, according to the training data set, a low-rank adaptation and thought chain combination technology is used to train a model to obtain a text review model, specifically including the following steps:

[0017] According to the training data set, a low-rank adaptation technology is used for low-rank adaptation training to obtain a preliminary model;

[0018] According to the power transformation review requirements, a thought chain template is constructed;

[0019] Based on the preliminary model and the thought chain template, combined training is performed to obtain a text review model.

[0020] By adopting the technical scheme, since substation text review involves complex professional judgment and reasoning process, such as the review of a maintenance scheme needs to consider multiple dimensions such as equipment parameters, operation sequence, safety measures, and the like, the traditional model training method often only focuses on the identification of a single feature, and it is difficult to realize deep logical reasoning; the application first performs efficient customization on the pre-trained model through low-rank adaptation technology, so that the pre-trained model obtains basic understanding ability in the substation field, then constructs a thinking chain template containing professional judgment logic based on the substation review requirements, and finally combines the two technologies organically to train the model; the low-rank adaptation ensures the professionalism of the model, and the thinking chain technology gives the model clear reasoning ability, which not only improves the review accuracy of the model, but also realizes the explainability of the review process.

[0021] Optionally, according to the training data set, low-rank adaptation training is performed using a low-rank adaptation technology to obtain a preliminary model, specifically including the following steps:

[0022] According to the training data set, the pre-trained model parameters are determined and frozen to obtain a basic model;

[0023] According to the basic model, a low-rank matrix is added in the transformer to obtain an extended model;

[0024] Based on the extended model, a bypass matrix is trained to obtain fine-tuning parameters;

[0025] According to the fine-tuning parameters, the pre-trained model parameters are updated to obtain a preliminary model.

[0026] By adopting the technical scheme, since the substation field is professional and special, it is difficult to accurately understand and process substation text content by directly using a general pre-trained model, and the traditional full-parameter fine-tuning method has the problems of high training cost and low efficiency; the application first freezes the pre-trained model parameters as a basic model, then adds a low-rank matrix in the transformer to construct an extended model, then obtains fine-tuning parameters by training a bypass matrix, and finally updates the model using these parameters; by introducing the low-rank matrix, the number of parameters that need to be trained is reduced, and the design of the bypass matrix ensures the rapid adaptation ability of the model to professional knowledge in the substation field; not only the calculation cost is greatly reduced, but also the adaptability of the model in the substation field is improved.

[0027] Optionally, according to the substation review requirements, a thinking chain template is constructed, specifically including the following steps:

[0028] Convert the substation review requirements into a machine-recognizable instruction format to obtain an instruction set;

[0029] Based on the instruction set, a standardized review step sequence is generated to obtain a reasoning flow;

[0030] Formalize the reasoning process to generate machine executable verification rules to obtain a rule set;

[0031] Convert the rule set into a structured mind chain prompt template to obtain a mind chain template.

[0032] By adopting the technical solution, the application first extracts and converts the audit requirements in various regulations into a standardized instruction format, then builds an audit step sequence with a clear sequence relationship based on the instructions, then converts the reasoning process into machine executable verification rules through formalization description, and finally generates a structured mind chain prompt template; not only the integrity and traceability of the audit logic are guaranteed, but also a flexible template expansion mechanism is provided, which provides a reliable reasoning basis for intelligent audit of substation text.

[0033] Optionally, based on the text audit model, a substation automation audit process is performed to obtain an audit result, specifically including the following steps:

[0034] Receiving a text to be audited, inputting the text to be audited into the text audit model to obtain a preliminary result;

[0035] Based on the preliminary result, a reasoning process is generated to obtain an audit record;

[0036] According to the audit record, an evaluation result and a suggestion are output to obtain an audit result.

[0037] By adopting the technical solution, the application first inputs the text to be audited into the audit model to obtain a preliminary judgment, then automatically generates a reasoning process containing key judgment nodes based on the output result of the model, and finally outputs an evaluation result and an improvement suggestion with guiding significance by analyzing the audit record; not only the transparency of the audit process is realized, but also a targeted improvement direction is provided, ensuring the reliability and practicality of the audit result.

[0038] Optionally, according to the audit record, an evaluation result and a suggestion are output to obtain an audit result, specifically including the following steps:

[0039] The use of substation terminology is evaluated, the accuracy of the terminology is calculated, and a terminology score is obtained;

[0040] The format of the substation text is evaluated, the format compliance is judged, and a format score is obtained;

[0041] The substation operation steps are evaluated, the step integrity is checked, and a completeness score is obtained;

[0042] Based on the terminology score, the format score and the completeness score, the safety of the substation is evaluated, the safety compliance is verified, and the audit result is output.

[0043] By adopting the technical scheme, the application firstly quantitatively evaluates the standardization of terms in the text through a professional term library, then verifies the standard degree of the text format according to a standard template, then checks the integrity and logicality of the operation steps based on a standard operation procedure, and finally integrates the scores of the three dimensions by weighting, and performs comprehensive safety evaluation; the quality of the power transformation text is quantified, and more reliable safety judgment is provided based on comprehensive analysis of multi-dimensional scores, which not only improves the accuracy and reliability of the evaluation result, but also provides clear direction guidance for text quality improvement.

[0044] In a second aspect, the application provides a power transformation text auditing system based on a large model, comprising:

[0045] A training data set acquisition module is configured to preprocess power transformation text materials to obtain a training data set;

[0046] A text auditing model acquisition module is configured to train a model by using a low-rank adaptation and thought chain combination technology according to the training data set to obtain a text auditing model;

[0047] An auditing result acquisition module is configured to execute a power transformation automation auditing process based on the text auditing model to obtain an auditing result.

[0048] In a third aspect, the application provides an electronic device comprising a memory, a processor, and a computer program stored on the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the power transformation text auditing method based on a large model.

[0049] In a fourth aspect, the application provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program is executed by a processor to implement the steps of the power transformation text auditing method based on a large model.

[0050] In summary, the application has at least one of the following beneficial technical effects:

[0051] 1. The application firstly preprocesses power transformation text materials to construct a training data set, then trains a model by using a low-rank adaptation and thought chain combination technology, and finally executes an automated auditing process to generate results; the accuracy and efficiency of text auditing are improved, and the application has strong expandability and can flexibly cope with changing business demands, effectively saving resources for manual review;

[0052] 2. The application first systematically collects substation regulations and operation manuals as original texts, constructs a professional dictionary system containing standard definitions of terms, usage specifications and scene associations, then performs term recognition and standardization based on the dictionary, and finally generates a training data set through tokenization and cleaning steps; the deep understanding and standardization of substation professional texts are realized, which not only improves the data quality, but also provides a reliable professional knowledge base for model training;

[0053] 3. Since substation text review involves complex professional judgment and reasoning process, such as the review of a maintenance scheme needs to consider multiple dimensions such as equipment parameters, operation sequence, safety measures, etc., the traditional model training method often only focuses on the recognition of single feature, and it is difficult to realize deep logical reasoning; the application first performs efficient customization on the pre-trained model through low-rank adaptation technology, so that it obtains the basic understanding ability in the substation field, then constructs a thinking chain template containing professional judgment logic based on the substation review requirements, and finally combines the two technologies for model training; the low-rank adaptation ensures the professionalism of the model, and the thinking chain technology gives the model clear reasoning ability, which not only improves the review accuracy of the model, but also realizes the explainability of the review process. BRIEF DESCRIPTION OF DRAWINGS

[0054] Figure 1 is a flowchart of a substation text review method based on a large model according to an embodiment of the application;

[0055] Figure 2 is a flowchart of step S100 in a substation text review method based on a large model according to an embodiment of the application;

[0056] Figure 3 is a flowchart of step S200 in a substation text review method based on a large model according to an embodiment of the application;

[0057] Figure 4 is a flowchart of step S210 in a substation text review method based on a large model according to an embodiment of the application;

[0058] Figure 5 is a flowchart of step S220 in a substation text review method based on a large model according to an embodiment of the application;

[0059] Figure 6 is a flowchart of step S300 in a substation text review method based on a large model according to an embodiment of the application;

[0060] Figure 7 is a flowchart of step S330 in a substation text review method based on a large model according to an embodiment of the application;

[0061] Figure 8 is a module schematic diagram of a power transformation text auditing system based on a large model according to an embodiment of the application;

[0062] Figure 9 is an internal structure diagram of an electronic device according to an embodiment of the application. DETAILED DESCRIPTION

[0063] The terms used in the following embodiments of the present application are only for the purpose of describing the specific embodiments and are not intended to be limiting of the present application. As used in the specification and the appended claims of the application, the singular forms "a," "an" and "the" are intended to include both singular and plural forms, unless the context clearly indicates otherwise. It will be further understood that the terms "and / or" as used herein refer to any or all possible combinations of one or more of the associated listed items.

[0064] Hereinafter, the terms "first", "second", "third", etc. are used only for the purpose of description and should not be understood as implying or suggesting relative importance or implicitly indicating the number of the technical features indicated. Therefore, the features defined with "first", "second" can explicitly or implicitly include one or more of the features, and in the description of the embodiments of the present application, the meaning of "a plurality of" is two or more, unless otherwise specified.

[0065] The embodiments of the present application will be further described in detail below with reference to the accompanying drawings.

[0066] In a first aspect, the present application provides a power transformation text auditing method based on a large model, referring to Figure 1 , comprising the following steps:

[0067] S100, pre-processing the power transformation text material to obtain a training data set.

[0068] The power transformation text material includes various types of professional documents such as substation operation rules, equipment operation manual, historical operation ticket, and maintenance record. These documents contain a large number of professional terms, standard operation procedures and safety specification requirements, which need to be systematically pre-processed before being used for model training. The purpose of pre-processing is to convert unstructured original text into standardized and normalized training data set.

[0069] Specifically, the text standardization processing technology is adopted, first, the document is uniformly formatted and the encoding is converted, then the professional field segmentation tool is used for segmentation processing, then the chapter structure and key information of the document are recognized through regular expression, and finally the processed text is organized into a training data set according to the predefined format. For example, for a substation switching operation ticket, the preprocessing process extracts the operation equipment name, operation steps, safety measures and other key information, and converts them into a structured data format.

[0070] S200, according to the training data set, a model is trained by using a low-rank adaptation and thought chain combination technology to obtain a text review model.

[0071] Among them, the model training adopts the low-rank adaptation and thought chain combination technology, which can quickly adapt to the professional knowledge in the substation field while maintaining the core ability of the pre-trained model. The low-rank adaptation technology realizes efficient field adaptation by adding a small number of trainable parameters in the model; the thought chain technology introduces a structured reasoning template, which enhances the logical reasoning ability of the model.

[0072] Specifically, first, a large-scale pre-trained model is used for low-rank adaptation training, and a training parameter matrix is added to capture the field characteristics; then, a thought chain reasoning template is constructed based on the review specifications in the substation field to guide the model to make standardized reasoning and judgment. For example, when reviewing a substation operation ticket, the model can perform step-by-step reasoning according to the thought chain of "equipment name verification -> operation step integrity check -> safety measure verification".

[0073] S300, based on the text review model, a substation automation review process is performed to obtain a review result.

[0074] Among them, the automation review process is a complete text analysis and evaluation process, including input text preprocessing, model reasoning, result evaluation and other links. The review model can comprehensively analyze the input substation text and finally output detailed review results and improvement suggestions.

[0075] Specifically, when receiving the text to be reviewed, the system first preprocesses and formats the text, and then inputs the processed text into the review model for analysis. The model will reason based on the trained parameters and thought chain templates to generate evaluation results in multiple dimensions. For example, the review result of a substation operation ticket will include term standardization score, step integrity score, safety measure sufficiency score, etc., and give specific problem explanations and improvement suggestions.

[0076] In one embodiment, referring to Figure 2 , the substation text materials include substation regulations and operation manuals; in step S100, the substation text materials are preprocessed to obtain a training data set, which specifically includes the following steps:

[0077] S110, read the substation regulation and operation manual files, and perform standardization processing to obtain structured raw text.

[0078] Among them, the substation regulation mainly includes equipment operation specification, safety operation requirement and accident handling process, and the operation manual includes operation step explanation, technical parameter and matters needing attention of specific equipment.

[0079] Specifically, first, use document parsing tools to convert files of different formats into pure text format, and then perform standardization processing such as unified encoding conversion, special character processing and line feed uniformity.

[0080] S120, based on the structured raw text, extract term features and construct a knowledge graph, establish a term mapping relationship table, and obtain term index data.

[0081] Among them, there are a large number of professional terms and technical terms in the substation field, and these terms have complex semantic association relationships. By constructing a knowledge graph, the hierarchical relationship and association relationship between terms can be clearly defined, which is helpful for subsequent term standardization processing. The nodes of the knowledge graph represent term concepts, and the edges represent the relationships between terms.

[0082] Specifically, professional terms are identified from structured text, including device names, operation actions, technical parameters, etc. Then, based on the co-occurrence relationship and context information of the terms, the semantic association between the terms is analyzed, and the knowledge graph is constructed.

[0083] S130, using term index data, performing term standardization identification and replacement processing on the text to generate standardized text.

[0084] Among them, different expressions of the same term often appear in substation text, which needs to be standardized based on term index data. This process includes term variant identification, standardization replacement and consistency check, etc. The purpose is to ensure that the use of terms in the text meets the standard requirements.

[0085] Specifically, the system first matches the term variant forms in the text using the term mapping relationship table, and then replaces the non-standard expressions identified with the corresponding standard terms. For example, different expressions such as "switch", "circuit breaker" and "switching device" in the text are uniformly standardized to the standard term "circuit breaker", ensuring the consistency of term usage.

[0086] S140, applying a word segmentation algorithm and data cleaning rules to the standardized text to generate a standard format training data set.

[0087] Among them, the standardized text needs to be further processed before it can be used for model training, which includes text segmentation, noise removal, format standardization and other steps.

[0088] Specifically, the text is segmented using a word segmentation dictionary optimized for the power transformation field, and stop words, special symbols and other irrelevant information in the text are removed. Then the processed text is organized according to the predefined data format, including the input text and the standard answer.

[0089] In one embodiment, referring to Figure 3 In step S200, a model is trained using low-rank adaptation and thought chain combination techniques based on the training data set to obtain a text review model, which includes the following steps:

[0090] S210, according to the training data set, using low-rank adaptation technology for low-rank adaptation training, obtaining a preliminary model.

[0091] Wherein, low-rank adaptation technology is a kind of efficient field adaptation method, through adding a small amount of trainable parameters in pre-training model to realize the directional adjustment of model ability. This method does not need to fine-tune the whole model, but captures the field knowledge by training specific low-rank matrix, which not only maintains the general ability of the model, but also realizes the rapid adaptation to the power transformation field.

[0092] Specifically, first select a suitable pre-training language model as the basis, then insert low-rank adaptation matrix between the key layers of the model. The rank of these matrices is much smaller than the dimension of the original parameter matrix, usually 5%-10% of the original dimension. For example, for a transformation matrix with a dimension of 1024, a low-rank adaptation matrix with a rank of 64 can be inserted. During training, only the parameters of these low-rank matrices are updated, while the other parameters of the pre-training model are fixed, so as to efficiently learn the feature representation of the power transformation field.

[0093] S220, according to the requirements of power transformation review, construct thought chain template.

[0094] Wherein, the thought chain template is a structured reasoning framework, which is used to guide the model to perform logical reasoning and judgment according to the predefined steps. In the power text review scene, according to the review requirements of different types of documents, the corresponding thought chain template needs to be constructed, including key check points, judgment standards and reasoning paths, etc.

[0095] S230, based on the preliminary model and the thought chain template, combination training is carried out to obtain a text review model.

[0096] Wherein, the combination training is the process of fusing the low-rank adapted preliminary model with the thought chain template, the purpose is to make the model can carry on the standardization of reasoning and judgment according to the predefined thought chain template.

[0097] Specifically, first, the training data is organized according to the structure of the thinking chain template, and a corresponding training sample is prepared for each reasoning step. Then, a loss function is designed, considering the matching degree of the model output and the standard answer, as well as the consistency of the reasoning process and the thinking chain template. For example, for the substation operation ticket review, the model needs to learn to first check the standardization of the equipment name, and then evaluate the integrity of the operation steps, the rationality of the operation sequence, etc., and finally give a comprehensive review result.

[0098] In one embodiment, with reference to Figure 4 In step S210, low-rank adaptation training is performed using low-rank adaptation techniques based on the training data set to obtain a preliminary model, including the following steps:

[0099] S211, according to the training data set, determine the pre-training model parameters and freeze, obtain the basic model.

[0100] The pre-training model is a language model trained on a large-scale general corpus, which has strong language understanding and generation capabilities. When applied to the substation field, the appropriate pre-training model and its parameter configuration need to be determined first, and these parameters are temporarily frozen to prevent damage to the original capabilities during subsequent training.

[0101] S212, according to the basic model, add low-rank matrices in the transformer to obtain an extended model.

[0102] The low-rank matrix is a parameter matrix with a small dimension that can effectively capture domain-specific features. These matrices are inserted into key positions in the transformer, including self-attention layers and feedforward network layers.

[0103] Specifically, two low-rank matrices are inserted into the attention calculation module of each transformer layer, which are used to adjust the calculation of the query vector and the key-value vector. The rank of the matrix is set to about 8% of the original dimension. For example, for a 768-dimensional hidden layer, a low-rank matrix with a rank of 60 can be inserted. These matrices are implemented by decomposing into two smaller matrices, i.e., the product of a 768x60 matrix and a 60x768 matrix.

[0104] S213, based on the extended model, train the bypass matrix to obtain fine-tuning parameters.

[0105] The bypass matrix refers to the additional trainable parameter matrix added outside the original model structure, which is parallel to the main calculation path of the model and is used to adjust the output of the model.

[0106] Specifically, a set of bypass matrices is designed for each transformer layer, including input mapping matrices and output mapping matrices. The dimensions of these matrices match the hidden state dimensions of the corresponding layers, but low-rank decomposition is used to reduce the number of parameters.

[0107] S214, updating the pre-training model parameters according to the fine-tuning parameters to obtain a preliminary model.

[0108] The fine-tuning parameters include the training results of the low-rank matrix and the bypass matrix, which need to be integrated with the original pre-training model parameters to form a unified model. The integration process needs to ensure that the new parameters can effectively affect the behavior of the model while not damaging the original language understanding ability of the model.

[0109] Specifically, the trained low-rank matrix and bypass matrix are combined with the original model parameters in a parameter combination manner. For each transformer layer, the effect of the low-rank matrix is added to the original calculation path through a residual connection. In this embodiment, the output of the original layer is h, the output of the low-rank matrix is h', and the final output is h+αh', where α is a learnable scaling factor.

[0110] In one embodiment, referring to Figure 5 In step S220, a thinking chain template is constructed according to the requirements of the substation audit, specifically including the following steps:

[0111] S221, converting the substation audit requirements into a machine-recognizable instruction format to obtain an instruction set.

[0112] S222, generating a standardized audit step sequence based on the instruction set to obtain a reasoning process.

[0113] The audit step sequence refers to organizing scattered audit instructions into a logically related execution sequence to form a complete reasoning process.

[0114] Specifically, the audit steps are organized in a directed acyclic graph manner, and steps with dependency relationships are connected by edges. For example, when auditing a substation operation ticket, a step sequence such as "device name verification -> operation environment inspection -> operation step verification -> safety measure confirmation" can be constructed.

[0115] S223, formally describing the reasoning process to generate machine-executable verification rules to obtain a rule set.

[0116] The formal description is the process of converting the reasoning process into strict logical expressions and judgment rules.

[0117] Specifically, conditional statements and logical operators are used to describe the verification rules.

[0118] S224, converting the rule set into a structured thinking chain prompt template to obtain a thinking chain template.

[0119] The thought chain prompting template is a structured text for guiding the model to perform reasoning, which converts the verification rules into a series of coherent reasoning steps.

[0120] Specifically, the template structure of "task description-reasoning step-judgment standard-output format" is adopted. For example, for operation ticket auditing, the template can include the following content: "Please audit the standardization of the following operation ticket. First step: check whether the equipment name is standardized, refer to the equipment naming specification... Second step: verify the completeness of the operation steps, confirm whether necessary safety measures are included...". Each step includes specific checking content and judgment standard, guiding the model to perform systematic reasoning analysis.

[0121] In one embodiment, referring to Figure 6 , in step S300, the substation automation auditing process is performed based on the text auditing model to obtain an auditing result, specifically including the following steps:

[0122] S310, receiving a text to be audited, inputting the text to be audited into the text auditing model to obtain a preliminary result.

[0123] S320, generating a reasoning process based on the preliminary result to obtain an auditing record.

[0124] S330, outputting an evaluation result and a suggestion according to the auditing record to obtain an auditing result.

[0125] In one embodiment, referring to Figure 7 , in step S330, the evaluation result and the suggestion are outputted according to the auditing record to obtain the auditing result, specifically including the following steps:

[0126] S331, evaluating the use of substation terminology, calculating a terminology accuracy to obtain a terminology score.

[0127] S332, evaluating the format of the substation text, judging a format compliance degree to obtain a format score.

[0128] S333, evaluating the substation operation steps, checking a step completeness degree to obtain a completeness score.

[0129] S334, performing a substation safety evaluation based on the terminology score, the format score and the completeness score, verifying a safety compliance degree, and outputting the auditing result.

[0130] It should be understood that the sequence of the steps in the above embodiments does not mean the order of execution, and the execution order of the processes should be determined according to their functions and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present application.

[0131] In a second aspect, the application provides a power transformation text auditing system based on a large model. The power transformation text auditing system based on a large model of the application is described below in combination with the power transformation text auditing method based on a large model.

[0132] With reference to Figure 8 A power transformation text auditing system based on a large model comprises:

[0133] A training data set acquisition module is configured to preprocess power transformation text materials to obtain a training data set.

[0134] A text auditing model acquisition module is configured to train a model using a low-rank adaptation and thought chain combination technique based on the training data set to obtain a text auditing model.

[0135] An auditing result acquisition module is configured to execute a power transformation automation auditing process based on the text auditing model to obtain an auditing result.

[0136] In an embodiment, the power transformation text materials comprise power transformation procedures and operation manuals. The training data set acquisition module comprises:

[0137] A text reading unit is configured to read power transformation procedure and operation manual files, perform standardization processing, and obtain structured original text.

[0138] A term extraction unit is configured to extract term features and construct a knowledge graph based on the structured original text, establish a term mapping relationship table, and obtain term index data.

[0139] A term processing unit is configured to perform term standardization identification and replacement processing on text using the term index data to generate standardized text.

[0140] A data cleaning unit is configured to apply a word segmentation algorithm and data cleaning rules to the standardized text to generate a standard format training data set.

[0141] In an embodiment, the text auditing model acquisition module comprises:

[0142] A low-rank adaptation training unit is configured to perform low-rank adaptation training using a low-rank adaptation technique based on the training data set to obtain a preliminary model.

[0143] A thought chain template construction unit is configured to construct a thought chain template based on power transformation auditing requirements.

[0144] A combination training unit is configured to perform combination training based on the preliminary model and the thought chain template to obtain a text auditing model.

[0145] In an embodiment, the low-rank adaptation training unit comprises:

[0146] The basic model obtaining subunit is configured to determine pre-training model parameters and freeze the parameters to obtain a basic model according to a training data set.

[0147] The extended model obtaining subunit is configured to add a low-rank matrix in the transformer to obtain an extended model according to the basic model.

[0148] The parameter training subunit is configured to train the bypass matrix based on the extended model to obtain fine-tuning parameters.

[0149] The model updating subunit is configured to update the pre-training model parameters according to the fine-tuning parameters to obtain a preliminary model.

[0150] In an embodiment, the thought chain template construction unit comprises:

[0151] The instruction conversion subunit is configured to convert the power transformation audit requirements into a machine-recognizable instruction format to obtain an instruction set.

[0152] The inference flow generation subunit is configured to generate a standardized audit step sequence based on the instruction set to obtain an inference flow.

[0153] The rule generation subunit is configured to formally describe the inference flow to generate a machine-executable verification rule to obtain a rule set.

[0154] The template conversion subunit is configured to convert the rule set into a structured thought chain prompt template to obtain a thought chain template.

[0155] In an embodiment, the audit result obtaining module comprises:

[0156] The preliminary result obtaining unit is configured to receive the text to be audited, input the text to be audited into the text audit model, and obtain a preliminary result.

[0157] The inference process generation unit is configured to generate an inference process based on the preliminary result to obtain an audit record.

[0158] The result output unit is configured to output an evaluation result and a suggestion according to the audit record to obtain an audit result.

[0159] In an embodiment, the result output unit comprises:

[0160] The term evaluation subunit is configured to evaluate the use of power transformation terms, calculate the term accuracy, and obtain a term score.

[0161] The format evaluation subunit is configured to evaluate the format of the power transformation text, judge the format compliance, and obtain a format score.

[0162] The step evaluation subunit is configured to evaluate the power transformation operation steps, check the step completeness, and obtain a completeness score.

[0163] The security evaluation subunit is configured to perform a power transformation security evaluation based on the term score, the format score and the integrity score, verify security compliance, and output an audit result.

[0164] In one embodiment, the present application provides an electronic device, which can be a server, and an internal structure diagram thereof can be as shown in Figure 9 The electronic device includes a processor, a memory and a network interface connected through a system bus. The processor of the electronic device is configured to provide computing and control capabilities. The memory of the electronic device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system, a computer program and a database. The internal memory provides an environment for running the operating system and the computer program in the non-volatile storage medium. The database of the electronic device is configured to store data. The network interface of the electronic device is configured to communicate with an external terminal through a network connection. The computer program is executed by the processor to implement a power transformation text auditing method based on a large model.

[0165] Those skilled in the art can understand that Figure 9 The structure shown in the above

[0166] In one embodiment, an electronic device is also provided, which includes a memory and a processor. The memory stores a computer program. The processor implements the steps in each of the above method embodiments when executing the computer program.

[0167] Those skilled in the art can understand that all or part of the processes in the above-mentioned embodiment methods can be completed by instructing the relevant hardware through a computer program. The above-mentioned computer program can be stored in a non-volatile computer readable storage medium, and when the computer program is executed, the computer program can include the processes of the above-mentioned embodiments of each method. Wherein, any reference to memory, storage, database or other medium used in each embodiment provided by the present application can include at least one of non-volatile and volatile memory. The non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory or optical memory, etc. The volatile memory can include random access memory (RAM) or external cache memory. As an illustration but not as a limitation, the RAM can be in various forms, such as static random access memory (SRAM) or dynamic random access memory (DRAM), etc.

[0168] The above are preferred embodiments of the present application, not to limit the protection scope of the present application, therefore: all equivalent changes made according to the structure, shape, principle of the present application should be covered within the protection scope of the present application.

Claims

1. A method for reviewing substation texts based on a large model, characterized in that, Includes the following steps: Read substation regulations and operation manuals, perform standardization processing, and obtain structured raw text; Based on the structured original text, terminology features are extracted and a knowledge graph is constructed. A terminology mapping relationship table is established to obtain terminology index data. The terminology mapping relationship table constructs hierarchical and relational relationships between terms based on the association relationships of professional terms in the substation field. Using the terminology index data, terminology standardization recognition and replacement processing is performed on the text to generate standardized text; The standardized text is then processed using word segmentation algorithms and data cleaning rules to generate a training dataset in a standard format. Based on the training dataset, low-rank adaptation training is performed using low-rank adaptation techniques to obtain a preliminary model; The substation audit requirements are converted into a machine-readable instruction format to obtain an instruction set; Based on the instruction set, a standardized sequence of review steps is generated to obtain the reasoning process; The reasoning process is formally described to generate machine-executable verification rules, thus obtaining a rule set. The rule set is converted into a structured thinking chain prompt template to obtain a thinking chain template. The thinking chain template adopts a template structure of "task description - reasoning steps - judgment criteria - output format". The corresponding thinking chain template is constructed in advance according to the review requirements of different types of documents in the substation text review scenario, including key checkpoints, judgment criteria and reasoning paths. The substation text review scenario includes the substation operation ticket review scenario. Based on the preliminary model and the thought chain template, a combined training is performed to obtain the text review model. The combined training is the process of fusing the preliminary model after low-rank adaptation with the thought chain template, with the aim of enabling the model to perform standardized reasoning and judgment according to the predefined thought chain template. Specifically, the text review model first organizes the training data according to the structure of the thought chain template and prepares corresponding training samples for each reasoning step. Then, a loss function is designed, taking into account the matching degree between the model output and the standard answer, as well as the consistency between the reasoning process and the thought chain template. Based on the text review model, the substation automation review process is executed to obtain the review results.

2. The substation text review method based on a large model according to claim 1, characterized in that, Based on the training dataset, low-rank adaptation training is performed using the low-rank adaptation technique to obtain a preliminary model, specifically including the following steps: Based on the training dataset, determine and freeze the parameters of the pre-trained model to obtain the base model; Based on the basic model, a low-rank matrix is ​​added to the transformer to obtain the extended model; Based on the extended model, the bypass matrix is ​​trained to obtain the fine-tuning parameters; The pre-trained model parameters are updated based on the fine-tuning parameters to obtain a preliminary model.

3. The substation text review method based on a large model according to claim 1, characterized in that, Based on the aforementioned text review model, a substation automation review process is executed to obtain the review results, specifically including the following steps: Receive the text to be reviewed, input the text to be reviewed into the text review model, and obtain preliminary results; Based on the preliminary results, a reasoning process is generated to obtain the audit record; Based on the audit records, output the evaluation results and suggestions to obtain the audit results.

4. The substation text review method based on a large model according to claim 3, characterized in that, Based on the audit records, the evaluation results and recommendations are output to obtain the audit results, which specifically include the following steps: The use of substation terminology is evaluated, the accuracy of the terminology is calculated, and a terminology score is obtained. The format of the substation text is evaluated to determine the degree of format conformity and obtain a format score; The substation operation procedures are evaluated to check their completeness and obtain a completeness score. Based on the terminology score, format score, and completeness score, a substation safety assessment is performed to verify safety compliance and output the audit results.

5. A substation text review system based on a large model, characterized in that, The substation text review method based on a large model, as described in any one of claims 1-4, includes: The training dataset acquisition module is used to preprocess the substation text data to obtain the training dataset; The text moderation model acquisition module is used to train the model using low-rank adaptation and mind chain combination techniques based on the training dataset to obtain the text moderation model. The audit result acquisition module is used to execute the substation automation audit process based on the text audit model and obtain the audit result.

6. An electronic device, characterized in that, It includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the substation text review method based on any one of claims 1-4.

7. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the steps of the substation text review method based on a large model as described in any one of claims 1-4.

Images