A large language model-based converter operation data analysis and early warning method

Abstract

The application discloses a method for analyzing and early warning of converter transformer operation and inspection data based on a large language model, and belongs to the technical field of electric power, and comprises the following steps: S1, constructing a converter transformer data analysis professional knowledge base; S2, analyzing converter transformer operation and inspection data based on the converter transformer data analysis professional knowledge base or a LLM large language model; and S3, based on the analysis result of the converter transformer data analysis professional knowledge base or the LLM large language model, if the analysis result is that the operation and inspection data of the converter transformer is abnormal, early warning is performed, otherwise, no early warning is performed; the application integrates automatic arrangement, analysis and early warning of converter transformer operation and inspection data, can not only assist operation and maintenance personnel in efficiently and accurately collecting data, but also can automatically analyze data to perform early warning, and improves the accuracy of analysis and early warning.

Description

Technical Field

[0001] The present invention belongs to the field of electric power technology, and in particular relates to a commutation transformer operation and inspection data analysis and early warning method based on a large language model. Background Art

[0002] With the rapid development of power systems, the scale and complexity of power grids are increasing, and the importance of safe operation of power grids is becoming increasingly prominent.

[0003] As the core equipment of the HVDC converter station, the operating status of the converter transformer is directly related to the safety and stability of the entire power grid. Therefore, daily operation and maintenance of the converter transformer is necessary. At the same time, the daily operation and maintenance data of the converter transformer also needs to be analyzed, including daily comparison, weekly analysis and monthly summary, to ensure that the operating status of the converter transformer can be understood in a timely manner and the operating status of the converter transformer can be maintained.

[0004] Currently, the analysis of daily operation and maintenance data of converter transformers, including daily comparison, weekly analysis, and monthly summary, is generally carried out by manually sorting, comparing, and analyzing large amounts of operation and maintenance data. This method has the following problems:

[0005] 1. Faced with massive amounts of data, manual processing is inefficient and time-consuming;

[0006] 2. Manual data processing is prone to errors, resulting in potential equipment failure risks not being discovered in time;

[0007] 3. A large amount of basic data processing work takes up the time of operation and maintenance personnel, preventing them from focusing on more important professional technical research and analysis.

[0008] To improve operation and maintenance efficiency and the accuracy of data analysis, the industry has proposed intelligent inspection systems, which use robots or drones to inspect equipment and collect data and image information. However, these systems mainly focus on physical inspections and are still unable to conduct in-depth data analysis.

[0009] In view of this, a commutator transformer operation and maintenance data analysis and early warning method based on a large language model is designed to solve the above problems. Summary of the Invention

[0010] To solve the problems raised in the above background technology, the present invention provides a converter transformer operation and inspection data analysis and early warning method based on a large language model. The method integrates the automatic organization, analysis and early warning of converter transformer operation and inspection data. It can not only assist operation and maintenance personnel in collecting data efficiently and accurately, but also automatically analyze the data for early warning, thereby improving the accuracy of analysis and early warning.

[0011] To achieve the above-mentioned object, the present invention provides the following technical solution: a method for analyzing and early warning commutation transformer operation and inspection data based on a large language model, comprising the following steps:

[0012] S1: Build a professional knowledge base for converter transformer data analysis;

[0013] S2: Analyze converter transformer operation and inspection data based on converter transformer data analysis expertise base or LLM large language model;

[0014] S3: Based on the analysis results of the converter transformer data analysis professional knowledge base or the LLM large language model, if the analysis result shows that the operation and inspection data of the converter transformer is abnormal, an early warning is issued; otherwise, no early warning is issued.

[0015] Furthermore, the specific steps in step S1 include:

[0016] S11: Collect converter transformer information, including technical documents, operation manuals, fault cases, and maintenance records;

[0017] S12: Preprocessing converter transformer data, including cleaning, denoising, format conversion, and key content annotation;

[0018] S13: Based on the LLM large language model, the pre-processed converter transformer data is inferred and segmented, divided into appropriate sizes, and vectorized and stored to form a professional knowledge base for converter transformer data analysis.

[0019] Furthermore, the specific steps of step S13 include:

[0020] S131: Load the specified converter transformer data, i.e., corpus, based on DirectoryLoader to form a document collection documents, and extract the content doc_text of the document collection documents;

[0021] S132: define the intelligent segmentation function split_text_with_llm of the LLM large language model, set the target text block size chunk_size of the intelligent segmentation function split_text_with_llm, the text segmentation prompt prompt, the minimum text block size min_chunk_size, the text block overlap size chunk_overlap, and the maximum recursion depth max_recursion_depth;

[0022] S133: The content doc_text is input into the LLM large language model, the content doc_text is inferred, and the size of the content doc_text is detected. If the length of the content doc_text is greater than the target text block size chunk_size, and the target text block size chunk_size is greater than the minimum text block size min_chunk_size, then the intelligent segmentation function split_text_with_llm is recursively called to split the content doc_text in half based on the text segmentation prompt prompt. The recursion depth is checked. If the current recursion depth current_depth is less than the maximum recursion depth max_recursion_depth, and the length of the segmented fragment of the content doc_text is greater than the target text block size chunk_size, and the target text block size chunk_size is greater than the minimum text block size min_chunk_size, then the above steps are repeated until the length of the segmented fragment of the content doc_text is less than or equal to the target text block size chunk_size. Then, the segmented fragment of the content doc_text is directly returned. Otherwise, the recursion is stopped and the segmented fragment of the content doc_text is directly returned.

[0023] S134: Encapsulate the segmented fragments into a Document object, including page_content and metadata, and generate a vector representation of the text block based on the embedding model;

[0024] S135: Create a storage table table_history_knowledge, including ID, keyword key_str and corresponding knowledge content fragment value_str, add the generated text block and embedding to the table table_history_knowledge, add the table table_history_knowledge to the vector database vectordb, and form a converter transformer data analysis professional knowledge base.

[0025] Furthermore, in step S131, after the content doc_text is extracted, the following operation steps are performed:

[0026] S1311: define a cutting function split_text_for_model, and set the text maximum context length max_context_length and text overlap rate overlap_percent of the cutting function split_text_for_model;

[0027] S1312: Split the content doc_text according to the set maximum context length max_context_length of the text based on the split_text_for_model cutting function to obtain the content doc_text that can be processed by the LLM large language model.

[0028] Furthermore, in step S1311, the maximum context length of the text max_context_length is 8000 characters, and the text overlap rate overlap_percent is 0.25.

[0029] Furthermore, in step S132, the target text block size chunk_size is 2048 characters, the text segmentation prompt prompt, the minimum text block size min_chunk_size is 512 characters, the text block overlap size chunk_overlap is half of the minimum text block size min_chunk_size, and the maximum recursion depth max_recursion_depth is 2.

[0030] Furthermore, the specific steps of step S2 include:

[0031] S21: establishing a connection with a converter transformer data analysis expertise database, including configuring a host, port, user name, and password;

[0032] S22: Based on the host input user's query input input_query_sentence;

[0033] S23: extract keywords from user input using regular expressions, and expand the extracted keywords into synonyms, antonyms, and professional terms, including Chinese and English expressions;

[0034] S24: Perform fuzzy matching query in the converter transformer data analysis professional knowledge base based on the expanded keywords;

[0035] S25: Set joined_value_str to an empty string, and store the corresponding knowledge content fragment value_str found in the joined_value_str;

[0036] S26: If all keywords are matched successfully, set all_matched=True and directly return the corresponding knowledge content fragment value_str in joined_value_str as the answer. Otherwise, set all_matched=False and call the LLM large language model;

[0037] S27: The LLM large language model performs big data retrieval analysis based on the user's query input input_query_sentence and expanded keywords.

[0038] Furthermore, in step S24, based on the fuzzy matching query analysis of the converter transformer operation and inspection data in the converter transformer data analysis professional knowledge base, an abnormality capture mechanism is set to capture and record abnormal information during the operation process for parameter optimization.

[0039] Furthermore, in step S27, during the process of big data retrieval and analysis of converter transformer operation and inspection data based on the LLM large language model, an abnormality capture mechanism is set to capture and record abnormal information during the operation process for parameter optimization.

[0040] Furthermore, in step S27, the big data retrieval and analysis results of the LLM large language model are further subjected to the following operation steps:

[0041] S271: Compare the big data retrieval and analysis results of the LLM large language model with the data in the converter transformer data analysis professional knowledge base, and extract new knowledge content from the big data retrieval and analysis results of the LLM large language model;

[0042] S272: Preprocessing new knowledge content, including length detection and filtering of invalid or erroneous results;

[0043] S273: extracting the keyword key_str from the new knowledge content using regular expressions;

[0044] S274: Compare the extracted keyword key_str with the keyword key_str in the converter transformer data analysis professional knowledge base. If the keyword key_str already exists in the converter transformer data analysis professional knowledge base, update the keyword key_str. If the keyword key_str does not exist in the converter transformer data analysis professional knowledge base, insert a new keyword key_str to implement the update of the converter transformer data analysis professional knowledge base.

[0045] Compared with the prior art, the present invention has the following beneficial effects:

[0046] 1. After collecting and preprocessing the data of the converter transformer, the present invention uses the LLM large language model for reasoning and segmentation to form a converter transformer data analysis professional knowledge base. Subsequently, the operation and inspection data of the converter transformer can be automatically sorted and analyzed by the converter transformer data analysis professional knowledge base. At the same time, automatic early warning can be issued based on the analysis results of the converter transformer data analysis professional knowledge base. That is, the automatic sorting, analysis and early warning of the converter transformer operation and inspection data are integrated. It can not only assist operation and maintenance personnel in collecting data efficiently and accurately, but also automatically analyze the data and issue early warnings, thereby improving the accuracy of analysis and early warning.

[0047] 2. The LLM large language model of the present invention performs multi-level segmentation on the converter transformer data through recursive calls, ensuring the size limit and semantic integrity of the segments, and improving the analysis accuracy of the converter transformer data analysis professional knowledge base.

[0048] 3. When the converter transformer data analysis professional knowledge base is unable to analyze the converter transformer operation and inspection data, the present invention can automatically call the LLM large language model for supplementary analysis, thereby improving the system's analysis accuracy of the converter transformer operation and inspection data.

[0049] 4. After calling the LLM large language model to perform supplementary analysis on the converter transformer operation and inspection data, the present invention automatically extracts new content from the LLM large language model analysis results for preprocessing and keyword extraction, and updates the keyword content of the converter transformer data analysis professional knowledge base. This can continuously expand the content of the converter transformer data analysis professional knowledge base and improve the comprehensiveness of the analysis of the converter transformer data analysis professional knowledge base. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] Figure 1 The flowchart of the converter transformer operation and inspection data analysis and early warning method based on the large language model of the present invention is shown. DETAILED DESCRIPTION

[0051] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0052] See also Figure 1 The present invention provides the following technical solution: a method for analyzing and early warning commutation transformer operation and inspection data based on a large language model, comprising the following steps:

[0053] S1: Build a professional knowledge base for converter transformer data analysis;

[0054] S11: Collect converter transformer information, including technical documents, operation manuals, fault cases, and maintenance records;

[0055] S12: Preprocessing converter transformer data, including cleaning, denoising, format conversion, and key content annotation;

[0056] Data cleaning and denoising: Automatically identify and process noise, outliers, and missing values ​​in data to improve data quality;

[0057] Format conversion: Convert data from different sources and formats into a unified standard format that can be processed by the LLM large language model;

[0058] S13: Based on the LLM large language model, the pre-processed converter transformer data is inferred and segmented, divided into appropriate sizes, and vectorized and stored to form a converter transformer data analysis expertise base;

[0059] S131: Load the specified converter transformer data, i.e., corpus, based on DirectoryLoader to form a document collection documents, and extract the content doc_text of the document collection documents;

[0060] After the content doc_text is extracted, the following steps are performed:

[0061] S1311: define a split_text_for_model cutting function, set the maximum context length of the text in the split_text_for_model cutting function max_context_length to 8000 characters, and the overlap rate of the text overlap_percent to 0.25;

[0062] S1312: Split the content doc_text according to the set maximum context length max_context_length based on the split_text_for_model function to obtain the content doc_text that can be processed by the LLM large language model;

[0063] S132: Define the intelligent segmentation function split_text_with_llm of the LLM large language model, set the target text block size chunk_size of the intelligent segmentation function split_text_with_llm to 2048 characters, set the text segmentation prompt prompt, the minimum text block size min_chunk_size to 512 characters, the text block overlap size chunk_overlap to half of the minimum text block size min_chunk_size, and the maximum recursion depth max_recursion_depth to 2;

[0064] The prompt is as follows:

[0065] prompt=f"""Please divide the following text into sections according to the following requirements:

[0066] 1. The size of the text fragment must be strictly limited! It must be less than {int(chunk_size)} characters;

[0067] 2. The content within the fragment should maintain semantic coherence. Do not split a paragraph in half, and do not split a table (including the notes to the table) in half.

[0068] 3. If a paragraph is too long and must be divided into multiple segments, each segment must be titled;

[0069] 4. If a table is too long and must be split into multiple segments, each segment must have a header;

[0070] 5. Add a subheading to each segment, no more than 100 words, on the first line, in the format: "###Title:XXXXXX###". The subheading should include the main keywords and a summary of the main content;

[0071] 6. Maintain semantic coherence within the paragraph as much as possible;

[0072] 7. If the size and semantic coherence requirements cannot be met simultaneously, add "### To be continued" at the end;

[0073] Please output in the following format:

[0074] - Each fragment begins with "Fragment {{number}}:" followed by the fragment content;

[0075] -Each fragment is separated by "===End of fragment===";

[0076] The following text needs to be segmented:

[0077] {text};

[0078] S133: The content doc_text is input into the LLM large language model, the content doc_text is inferred, and the size of the content doc_text is detected. If the length of the content doc_text is greater than the target text block size chunk_size, and the target text block size chunk_size is greater than the minimum text block size min_chunk_size, then the intelligent segmentation function split_text_with_llm is recursively called to split the content doc_text in half based on the text segmentation prompt prompt and the segmentation format of "=== segment end ===". The recursion depth is checked. If the current recursion depth current_depth is less than the maximum recursion depth max_recursion_depth, and the length of the segmented fragment of the content doc_text is greater than the target text block size chunk_size, and the target text block size chunk_size is greater than the minimum text block size min_chunk_size, then the above steps are repeated until the length of the segmented fragment of the content doc_text is less than or equal to the target text block size chunk_size. The segmented fragments of the content doc_text are directly returned as streaming output. Otherwise, the recursion is stopped and the segmented fragments of the content doc_text are directly returned as streaming output.

[0079] S134: Encapsulate the segmented fragments into a Document object, including page_content and metadata, and generate a vector representation of the text block based on the embedding model;

[0080] S135: creating a storage table table_history_knowledge, including an ID, a keyword key_str, and a corresponding knowledge content fragment value_str, adding the generated text block and embedding to the table table_history_knowledge, and adding the table table_history_knowledge to the vector database vectordb, thereby forming a converter transformer data analysis professional knowledge base;

[0081] S2: Analyze converter transformer operation and inspection data based on converter transformer data analysis expertise base or LLM large language model;

[0082] S21: establishing a connection with a converter transformer data analysis expertise database, including configuring a host, port, user name, and password;

[0083] S22: Based on the host input user's query input input_query_sentence;

[0084] The query input input_query_sentence of the user, i.e., real-time operation and inspection data, includes device operation parameters, environmental data, historical fault records, etc.

[0085] S23: Key words are extracted from the user input by using a regular expression, and the extracted key words are expanded for synonyms, near-synonyms, and professional terms, including Chinese expressions and English expressions;

[0086] S24: Based on the expanded key words, fuzzy matching queries are performed in the converter transformer data analysis professional knowledge base;

[0087] During the process of analyzing the converter transformer operation and inspection data based on fuzzy matching queries in the converter transformer data analysis professional knowledge base, an exception capture mechanism is set to capture and record abnormal information in the operation process, so as to optimize parameters;

[0088] S25: Set joined_value_str to be an empty string, and store the corresponding knowledge content segment value_str found in joined_value_str;

[0089] S26: If all key words are successfully matched, set all_matched=True, and directly return the corresponding knowledge content segment value_str in joined_value_str as an answer, otherwise, set all_matched=False, and call the LLM large language model;

[0090] S27: The LLM large language model performs big data retrieval analysis based on the query input input_query_sentence of the user and the expanded key words;

[0091] During the process of analyzing the converter transformer operation and inspection data based on the big data retrieval analysis of the LLM large language model, an exception capture mechanism is set to capture and record abnormal information in the operation process, so as to optimize parameters;

[0092] Based on the analysis results of the converter transformer data analysis professional knowledge base or the LLM large language model, requirements raised by the user are displayed in the form of intuitive line graphs, curve graphs, scatter graphs, histogram graphs, and pie charts;

[0093] The big data retrieval analysis results of the LLM large language model also perform the following operation steps:

[0094] S271: Compare the big data retrieval analysis results of the LLM large language model with the data in the converter transformer data analysis professional knowledge base, and extract new knowledge content in the big data retrieval analysis results of the LLM large language model;

[0095] S272: Preprocessing new knowledge content, including length detection and filtering of invalid or erroneous results;

[0096] S273: extracting the keyword key_str from the new knowledge content using regular expressions;

[0097] S274: comparing the extracted keyword key_str with the keyword key_str in the converter transformer data analysis professional knowledge base; if the keyword key_str already exists in the converter transformer data analysis professional knowledge base, updating the keyword key_str; if the keyword key_str does not exist in the converter transformer data analysis professional knowledge base, inserting a new keyword key_str to update the converter transformer data analysis professional knowledge base;

[0098] S3: Based on the analysis results of the converter transformer data analysis professional knowledge base or the LLM large language model, if the analysis result shows that the operation and inspection data of the converter transformer is abnormal, an early warning is issued; otherwise, no early warning is issued.

[0099] While embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions, and variations may be made to these embodiments without departing from the principles and spirit of the invention, and that the scope of the invention is defined by the appended claims and their equivalents.

Claims

1. A method for analyzing and warning commutation transformer operation and inspection data based on a large language model, characterized in that: The following steps are involved: S1: Build a professional knowledge base for converter transformer data analysis; S2: Analyze converter transformer operation and inspection data based on converter transformer data analysis expertise base or LLM large language model; S3: Based on the analysis results of the converter transformer data analysis professional knowledge base or the LLM large language model, if the analysis result shows that the operation and inspection data of the converter transformer is abnormal, an early warning is issued; otherwise, no early warning is issued; The specific steps in step S1 include: S11: Collect converter transformer information, including technical documents, operation manuals, fault cases, and maintenance records; S12: Preprocessing converter transformer data, including cleaning, denoising, format conversion, and key content annotation; S13: Based on the LLM large language model, the pre-processed converter transformer data is inferred and segmented, divided into appropriate sizes, and vectorized and stored to form a converter transformer data analysis expertise base; The specific steps of step S13 include: S131: Load the specified converter transformer data, i.e., corpus, based on DirectoryLoader to form a document collection documents, and extract the content doc_text of the document collection documents; S132: define the intelligent segmentation function split_text_with_llm of the LLM large language model, set the target text block size chunk_size of the intelligent segmentation function split_text_with_llm, the text segmentation prompt prompt, the minimum text block size min_chunk_size, the text block overlap size chunk_overlap, and the maximum recursion depth max_recursion_depth; S133: The content doc_text is input into the LLM large language model, the content doc_text is inferred, and the size of the content doc_text is detected. If the length of the content doc_text is greater than the target text block size chunk_size, and the target text block size chunk_size is greater than the minimum text block size min_chunk_size, then the intelligent segmentation function split_text_with_llm is recursively called to split the content doc_text in half based on the text segmentation prompt prompt. The recursion depth is checked. If the current recursion depth current_depth is less than the maximum recursion depth max_recursion_depth, and the length of the segmented fragment of the content doc_text is greater than the target text block size chunk_size, and the target text block size chunk_size is greater than the minimum text block size min_chunk_size, then the above steps are repeated until the length of the segmented fragment of the content doc_text is less than or equal to the target text block size chunk_size. Then, the segmented fragment of the content doc_text is directly returned. Otherwise, the recursion is stopped and the segmented fragment of the content doc_text is directly returned. S134: Encapsulate the segmented fragments into a Document object, including page_content and metadata, and generate a vector representation of the text block based on the embedding model; S135: Create a storage table table_history_knowledge, including ID, keyword key_str and corresponding knowledge content fragment value_str, add the generated text block and embedding to the table table_history_knowledge, add the table table_history_knowledge to the vector database vectordb, and form a converter transformer data analysis professional knowledge base.

2. The method for analyzing and early warning commutation transformer operation and inspection data based on a large language model according to claim 1, characterized in that: In step S131, after the content doc_text is extracted, the following steps are performed: S1311: define a cutting function split_text_for_model, and set the maximum context length max_context_length and the text overlap rate overlap_percent of the cutting function split_text_for_model; S1312: Split the content doc_text according to the set maximum context length max_context_length of the text based on the split_text_for_model cutting function to obtain the content doc_text that can be processed by the LLM large language model.

3. The method for analyzing and early warning commutation transformer operation and inspection data based on a large language model according to claim 2, characterized in that: In step S1311 , the maximum context length of the text max_context_length is 8000 characters, and the text overlap rate overlap_percent is 0.

25.

4. The method for analyzing and early warning commutation transformer operation and inspection data based on a large language model according to claim 1 is characterized by: In step S132 , the target text block size chunk_size is 2048 characters, the text segmentation prompt prompt is used, the minimum text block size min_chunk_size is 512 characters, the text block overlap size chunk_overlap is half of the minimum text block size min_chunk_size, and the maximum recursion depth max_recursion_depth is 2.

5. The method for analyzing and early warning commutation transformer operation and inspection data based on a large language model according to claim 1 is characterized by: The specific steps of step S2 include: S21: establishing a connection with a converter transformer data analysis expertise database, including configuring a host, port, user name, and password; S22: Based on the host input user's query input input_query_sentence; S23: extract keywords from user input using regular expressions, and expand the extracted keywords into synonyms, antonyms, and professional terms, including Chinese and English expressions; S24: Perform fuzzy matching query in the converter transformer data analysis professional knowledge base based on the expanded keywords; S25: Set joined_value_str to an empty string, and store the corresponding knowledge content fragment value_str found in joined_value_str; S26: If all keywords are matched successfully, set all_matched=True and directly return the corresponding knowledge content fragment value_str in joined_value_str as the answer. Otherwise, set all_matched=False and call the LLM large language model; S27: The LLM large language model performs big data retrieval analysis based on the user's query input input_query_sentence and expanded keywords.

6. The method for analyzing and early warning commutation transformer operation and inspection data based on a large language model according to claim 5 is characterized by: In step S24, based on the fuzzy matching query analysis of the converter transformer operation and inspection data in the converter transformer data analysis professional knowledge base, an abnormality capture mechanism is set to capture and record abnormal information during the operation process for parameter optimization.

7. The method for analyzing and early warning commutation transformer operation and inspection data based on a large language model according to claim 5 is characterized by: In step S27, during the process of big data retrieval and analysis of converter transformer operation and inspection data based on the LLM large language model, an abnormality capture mechanism is set to capture and record abnormal information during the operation process in order to perform parameter optimization.

8. The method for analyzing and early warning commutation transformer operation and inspection data based on a large language model according to claim 5 is characterized by: In step S27, the big data retrieval and analysis results of the LLM large language model further undergo the following operation steps: S271: Compare the big data retrieval and analysis results of the LLM large language model with the data in the converter transformer data analysis professional knowledge base, and extract new knowledge content from the big data retrieval and analysis results of the LLM large language model; S272: Preprocessing new knowledge content, including length detection and filtering of invalid or erroneous results; S273: extracting the keyword key_str from the new knowledge content using a regular expression; S274: Compare the extracted keyword key_str with the keyword key_str in the converter transformer data analysis professional knowledge base. If the keyword key_str already exists in the converter transformer data analysis professional knowledge base, update the keyword key_str. If the keyword key_str does not exist in the converter transformer data analysis professional knowledge base, insert a new keyword key_str to implement the update of the converter transformer data analysis professional knowledge base.

Images