Large-model-based structured extraction method and system for electric power overhaul data

Abstract

The invention relates to a structured extraction method and system for electric power overhaul data based on a large model. The method comprises the following steps: collecting and processing original electric power overhaul data; loading and analyzing the standardized naming mapping table of the power system equipment; configuring and calling a large language model, and performing feature extraction on the preprocessed to-be-analyzed text to obtain structured information of a target field; performing semantic similarity calculation based on vector embedding to obtain a corresponding preliminary matching result; further calculating the preliminary matching result according to a standardization algorithm, and establishing a standardization mapping index table of the current plant station and each level of equipment in the station; and the standardized mapping index table is integrated and packaged into a system service or offline file form to be output. According to the method, efficient and accurate structured extraction and standardized processing of electric power overhaul data are realized, complex unstructured text input can be processed, and extracted information is accurately corresponding to standardized equipment and station names through semantic understanding and a multi-stage matching mechanism. The process greatly improves the automation degree and accuracy of data processing.

Description

Technical Field

[0001] The present invention relates to the technical field of power system data processing, and in particular to a structured extraction method and system for power maintenance data based on a large model. Background Art

[0002] In the daily operation of the power system, the accuracy, consistency and traceability of maintenance data are crucial to ensure the safe and stable operation of the power grid. However, a common problem is that the filling of maintenance content mainly relies on manual operation, which leads to many challenges in data quality and consistency:

[0003] 1. Data inconsistency: Due to differences in personal habits, professional backgrounds and expressions of the reporting personnel, even the maintenance content of the same equipment may be described in different ways. For example, the same transformer may be described as "XX transformer", "XX main transformer" or customary abbreviations.

[0004] 2. Lack of standardization: Although the power equipment in the power system has standardized naming, because there are many devices involved in maintenance and the operation is complicated, if the data is reported by equipment selection in the maintenance system, it will increase the complexity of operation and affect the operability of the system. Therefore, strict data standardization is often not implemented in the current maintenance system. The reporting personnel may use non-standard abbreviations, common names or personal expressions, which makes subsequent data processing and analysis complicated.

[0005] 3. Information islands between systems: Due to the lack of unified naming conventions and data standards, data between different systems are difficult to effectively link and share, forming information islands.

[0006] 4. Difficulty in closing the business loop: Data inconsistency and non-standardization seriously hinder the closing of cross-system business processes. For example, it is difficult to accurately track the complete maintenance history of a specific device, or to accurately match maintenance records with equipment information in the asset management system.

[0007] 5. Limited data analysis: Unstructured and non-standardized data formats greatly limit the analytical value of data. This not only increases the workload of data preprocessing, but may also lead to deviations or errors in analysis results.

[0008] 6. Heavy burden of manual processing: Due to data inconsistency, a lot of manual intervention and data cleaning work becomes necessary, which is not only time-consuming and labor-intensive, but also prone to introduce new errors.

[0009] 7. Insufficient decision support capabilities: Data quality issues directly affect the effectiveness of the decision support system based on these data, which may lead to improper resource allocation or unreasonable maintenance strategy formulation.

[0010] The above problems not only affect the efficiency of daily operation and maintenance, but also pose potential risks to the safe and stable operation of the power grid. As the power system develops towards intelligence and digitalization, the requirements for data quality are getting higher and higher. Traditional manual reporting and simple data processing methods can no longer meet the needs of modern power systems.

[0011] Therefore, there is an urgent need for a method and system that can automatically process non-standardized maintenance data and realize data structuring and standardized extraction. This method and system should be able to understand and process the diverse manually filled content, extract key information, and convert it into a standardized naming and structured format, thereby supporting cross-system data integration and business closed loop, and improving data availability and analytical value. Summary of the invention

[0012] In view of the deficiencies in the prior art, the present invention utilizes the powerful context understanding capability of the large language model to provide a structured extraction method and system for power maintenance data based on the large model to solve the above-mentioned problems existing in the prior art.

[0013] The technical solution adopted by the present invention to achieve the above-mentioned purpose is: a structured extraction system of power maintenance data based on a large model, which constructs the following system for accurately matching the extracted information with standardized equipment and plant names through semantic understanding and multi-level matching mechanism for multi-source complex heterogeneous texts, and provides them to maintenance users. The system includes:

[0014] Data acquisition preprocessing module, collecting raw power maintenance data, data cleaning and formatting;

[0015] A standardized mapping table preparation module is used to load and parse the standardized naming mapping table of power system equipment;

[0016] The unstructured text feature extraction module configures and calls the large language model to extract features from the preprocessed text to be analyzed and obtain the structured information of the target field;

[0017] The semantic similarity module uses semantic similarity calculation based on vector embedding to match the structured information of the target field and search the standardized naming mapping table to obtain the corresponding preliminary matching results;

[0018] The plant and equipment name standardization module further calculates the preliminary matching results according to the standardization algorithm, establishes the standardized mapping index table of the current plant and each level of equipment in the plant, and realizes data standardization;

[0019] The data integration and output module integrates the standardized mapping index table, encapsulates it into a system service or offline file, and outputs it to the power grid business system for use.

[0020] The power maintenance data is multi-source heterogeneous text records, Excel tables, and database data;

[0021] The data acquisition preprocessing module includes:

[0022] The secure transmission unit uses a secure file transfer protocol to obtain data files from a remote server; it is used to prevent eavesdropping, tampering or forgery during the transmission process;

[0023] A local cache unit, used for local cache of data files;

[0024] The data parsing unit parses the multi-source heterogeneous data into a unified data format; including: using regular expressions to filter out a number of target data that meet the conditions in the data set, splitting or searching the target data according to the specified characters, and obtaining a set of segmented target data;

[0025] The data cleaning unit is used to implement data cleaning, including removing irrelevant characters, unifying separators, and format standardization.

[0026] The standardized mapping table preparation module includes:

[0027] The multi-level mapping structure unit is used to collect the original mapping table, and parse and process the ID mapping data of the corresponding level into: plant station mapping table and equipment mapping table; it is used to meet the needs of different scenarios; wherein the original mapping table is a standardized naming mapping table of existing power system equipment; wherein each standardized naming mapping table stores the standard naming names of all plants and equipment under the current system key value without hierarchy; the parsing operation is to parse each entry of the original mapping table of the current plant station into the plant station mapping table and the equipment mapping table according to the hierarchy;

[0028] A mapping data loading unit, used for reading ID mapping data from a text file;

[0029] A rule unit is used to define the consistency of the naming rules of each standardized name in the ID mapping data and the data to be analyzed; and to match the mapping key value of each standardized name in the ID mapping data with the data to be analyzed; wherein the mapping key value is a classification identifier of data from different sources, and the consistent naming rules are used to match data from different sources.

[0030] The unstructured text feature extraction module includes:

[0031] The large language model unit is used to adopt the large language model and combine it with the parameter unit to iteratively train the input data to obtain an ideal model, so as to extract the structured information of the target field; the key target fields include: XX transformer, XX transformer substation, XX plant, XX power plant, XX wind farm, and XX wind farm;

[0032] The parameter unit is used to define the roles, tasks, instruction templates, output data structures, and domain knowledge of the large model; the tasks, instruction templates, output data structures, and domain knowledge under different roles are different; the domain knowledge inputs common equipment types and maintenance terms in the power system into the model to help the model learn domain-specific expressions.

[0033] The semantic similarity module includes:

[0034] The semantic similarity calculation unit implements semantic similarity calculation based on vector embedding, which is used for matching and searching the structured information of the plant and equipment names in the text to be extracted with the standardized names in the plant and equipment mapping tables parsed from different system sources.

[0035] A text vector representation unit is used to represent a given text t vector as: V(t)=f(t); wherein t is the structured information of the plant station and equipment name in the text to be processed or the standardized name in the plant station mapping table and equipment mapping table after parsing under different system sources, and f is an encoding function;

[0036] The similarity calculation unit is used to calculate the similarity and search and match the most similar items in different ranges.

[0037] The similarity calculation includes: for the query text q and the text set T = {t1, t2, ..., tn}, calculating the similarity of each item according to the following formula:

[0038]

[0039] Among them, the query text q is the structured information of the plant and equipment names in the text to be processed, and the text set T is a large set of standardized names in the plant mapping table and equipment mapping table parsed from different system sources, where · represents the dot product, and |V| represents the Euclidean norm of the vector.

[0040] The searching and matching of the most similar items in different ranges includes:

[0041] Select the most similar item in set T as the matching result according to the following formula;

[0042] Calculate most_similar(q, T) = argmax t∈T (similarity(q,t));

[0043] Alternatively, the similarity values ​​are arranged in descending order, and the first M are taken as the candidate name set c = {c1, c2, ..., cn} as the matching result output.

[0044] The standardization algorithm of the plant and equipment name standardization module is used to handle various naming variants and abbreviations; including:

[0045] a) Edit distance calculation: For two strings s1 and s2, calculate the edit distance D(i, j):

[0046]

[0047] Wherein, the string s1 is the structured information of the plant and equipment names in the text to be processed, s2 is any one of the candidate names c = {c1, c2, ..., cn}; D(i, j) is the semantic similarity distance;

[0048] b) For candidate name c and standard name s, semantic similarity and edit distance are comprehensively considered to calculate the standardized score:

[0049] score(c,s)=α·similarity(c,s)+(1-α)·(1-normalized_edit_distance(c,s))

[0050] Among them, α is the weight parameter, similarity(c,s) is the result of calculating c and s according to the similarity formula, and normalized_edit_distance is the normalized edit distance D(i, j);

[0051] c) For a candidate name c and a standard name set s, the selection of the best match is expressed as:

[0052] best_match(c, S) = argmax s∈S score(c,s)

[0053] At this point, a unique standardized mapping index table of the target fields and standard names of all plant and equipment names is established.

[0054] The data integration and output module uses the JSON serialization method to format the data in the standardized mapping index table; uses an incremental data processing and output mechanism to support efficient processing and real-time output of large-scale data; is used to implement the data upload function and securely transmit the processing results back to the remote server for use by the maintenance system.

[0055] A structured extraction method for power maintenance data based on a large model performs the following steps to identify multi-source complex and heterogeneous power maintenance data, performs matching retrieval in a standardized mapping table to obtain a unique standard name and provide it to maintenance users. The method includes the following steps:

[0056] Step S1, collecting original power maintenance data, cleaning and formatting the data;

[0057] Step S2, loading and parsing the standardized naming mapping table of power system equipment;

[0058] Step S3: configure and call a large language model to extract features from the preprocessed text to be analyzed to obtain structured information of the target field;

[0059] Step S4: using semantic similarity calculation based on vector embedding to match the structured information of the target field and search the standardized naming mapping table to obtain the corresponding preliminary matching result;

[0060] Step S5: further calculate the preliminary matching results according to the standardized algorithm, establish a standardized mapping index table for the current plant and each level of equipment in the plant, and realize data standardization;

[0061] Step S6: Integrate the standardized mapping index table, encapsulate it into a system service or an offline file, and output it to the power grid business system for use.

[0062] Compared with the prior art, the present invention has the following beneficial effects and advantages:

[0063] The present invention realizes efficient and accurate structured extraction and standardized processing of power maintenance data through the above detailed technical steps and method designs. The method can process complex unstructured text inputs and accurately match the extracted information with standardized equipment and plant names through semantic understanding and multi-level matching mechanisms. This process greatly improves the automation and accuracy of data processing, laying a solid foundation for subsequent data analysis and business applications. BRIEF DESCRIPTION OF THE DRAWINGS

[0064] Figure 1 It is a schematic diagram of the system flow of the present invention;

[0065] Figure 2 It is a system structure diagram of the present invention. DETAILED DESCRIPTION

[0066] In order to make the above-mentioned purposes, features and advantages of the present invention more obvious and easy to understand, the specific implementation method of the present invention is further described in detail below in conjunction with the accompanying drawings. In the following description, many specific details are set forth to facilitate a full understanding of the present invention, but the present invention can be implemented in many other ways different from those described herein, and those skilled in the art can make similar improvements without violating the connotation of the invention, so the present invention is not limited by the specific implementation disclosed below.

[0067] like Figure 1-2 As shown, a structured extraction method of power maintenance data based on a large model includes:

[0068] 1. Data acquisition and preprocessing: Obtain raw power maintenance data from multiple sources, including but not limited to text records, Excel tables, databases, etc.; the data acquisition format is as follows:

[0069] B4XXXXX-8XXX-AXXX-BXXX-B5XXXXXXXXXX@@@XX.XX plant, unit 5, spot check, handling of X-watt vibration defect of unit X. @@@ shutdown.

[0070] 7DXXXXX-5XXX-9XXX-EXXX-CAXXXXXXXX@@@XXX factory, machine No. 1, standby@@@ shut down.

[0071] 2. Standardized mapping table preparation: Establish and load the standardized naming mapping table for power system equipment to provide a basis for subsequent data standardization;

[0072] 3. Extraction from unstructured text: Configure and call a large language model to extract structured information of the target field from the preprocessed text to be analyzed;

[0073] 4. Semantic similarity calculation: implement semantic similarity calculation based on vector embedding, which is used to extract the target field (power system equipment name) in the text to be extracted and match the standardized name (in the power system equipment standardized naming mapping table) to determine the standardized name content corresponding to the target field.

[0074] 5. Standardization of plant and equipment names: Match the extracted target fields (equipment and plant names) with the standardized mapping table one by one, establish a standardized mapping index table for the current plant and each level of equipment within the plant, and achieve data standardization.

[0075] 6. Data integration and output: Integrate the standardized structured data (in the standardized mapping index table) and output it in the form of system services or offline files for use by the power grid business system.

[0076] The specific implementation steps of the technical solution are as follows:

[0077] Step 1: Data acquisition and preprocessing: obtain raw power maintenance data from multiple sources and perform preliminary cleaning and format unification.

[0078] Step 101: Design a secure file transfer mechanism to support secure acquisition of data files from a remote server. The secure file transfer mechanism uses the commonly used SFTP file transfer mechanism, which is a secure protocol for file transfer via the SSH protocol. It uses the encryption function of SSH to encrypt and transmit files, ensuring that files are not eavesdropped, tampered with, or forged during transmission, and protecting the confidentiality and integrity of files.

[0079] Step 102: Implement a local cache strategy for data files to improve processing efficiency and support offline operations.

[0080] Step 103: A general data parsing algorithm is used to process input data of different formats, such as text data separated by delimiters. The data parsing algorithm needs to find and parse data of the following format, for example: B4XXXXX-8XXX-AXXX-BXXX-B5XXXXXXXXXX@@@XX.XX plant, unit 5, spot check, processing of large vibration defect of unit A B tile. @@@ shutdown.

[0081] You can first use regular expressions to filter out several target data that meet the conditions in the data set, and then split or search the target data according to the specified characters. Finally, you get a set of segmented target data:

[0082] B4XXXXX-8XXX-AXXX-BXXX-B5XXXXXXXXXX, XX.XX factory, No. 5 machine.

[0083] Step 104: Implement data cleaning functions, including removing irrelevant characters, unifying separators, and format standardization, such as processing XX#1 line into XX No. 1 line.

[0084] Step 2: Prepare the standardized mapping table, establish and load the standardized naming mapping table for power system equipment, and provide a basis for subsequent data standardization.

[0085] Step 201: Design a multi-level ID mapping structure, including an original mapping table, and disassemble it into ID mapping data: a plant mapping table and an equipment mapping table, to meet the needs of different scenarios. Among them, the original mapping table is an existing standardized naming mapping table for power system equipment (for example, data from different systems, and multiple standardized naming mapping tables exist for multiple systems), wherein each standardized naming mapping table stores all the standard naming names of the relevant plants and equipment under the current system key value without hierarchy; the post-processing mapping operation is to disassemble the various entries of the original mapping table of the current plant into a plant mapping table and an equipment mapping table according to the plant, equipment and other levels. The hierarchy is exemplified as follows: the first level includes plant one, plant two...; the second level includes the power transmission and transformation equipment under each plant...; the third level includes the sensor equipment, switches and other actuators under each power transmission and transformation equipment...

[0086] For example, the original mapping table may be in the following form:

[0087] 99XXXXXXXXXXX@@@XX factory@@@1

[0088] 99XXXXXXXXXXX@@@XX Factory@@@99XXXXXXXXXX1@@@Unit No.1

[0089] The processed mapping table is in the form of:

[0090] {

[0091] "id":"XXXXXXXXX",

[0092] "station":{"confirm_name":"XX factory","confirm_id":"99XXXXXXXXXXX"},

[0093] "devices":[{confirm_name":machine No. 1","confirm_id":"99XXXXXXXXXX1"}]

[0094] }

[0095] Step 202: Develop a mapping data loading algorithm that can read ID mapping data from a text file.

[0096] Step 203: Implement a special character processing function in the standardized name to make the standardized name consistent with the naming rules of the data to be analyzed.

[0097] Step 204: Design a unified mapping key processing method to ensure that data from different sources can be matched consistently. The mapping key is a classification identifier for data from different sources.

[0098] Step 3: Extract from unstructured text, configure and call a large language model to extract structured information from the text to be analyzed.

[0099] Step 301: Design a model initialization and warm-up mechanism to ensure that the model is in the best condition before use, reduce the inference time of large models, and improve the response speed.

[0100] Step 302: Design a dedicated instruction template to guide the large model to accurately understand and extract key information from the power maintenance data. Including:

[0101] 1. Assign roles to the big model. For example, tell the big model the following instructions: You are now an electric power operation and maintenance engineer, and you need to identify the equipment to be identified according to the text given to you and the following requirements.

[0102] 2. Clearly define the task requirements: clearly indicate the key information (target fields) that need to be extracted from the text, such as equipment names, etc. Target fields include: XX transformer, XX substation, XX plant, XX power plant, XX wind farm, XX wind farm

[0103] 3. Output format requirements: Specify the output JSON structure, including the name and data type of each target field.

[0104] 4. Domain knowledge prompts: Input key information such as common equipment types and maintenance terms in the power system into the model to help the model understand domain-specific expressions.

[0105] 5. Example description: Give an example of input text (target field) and expected output in JSON format to help the model understand the task requirements. An example of structured information in JSON format is as follows: {station_name: "power plant or substation name", device: ["device name 1", "device name 2"]}

[0106] Step 303: Implement the large model API calling interface, including request construction, sending and response processing.

[0107] Step 304: Design a JSON data extraction algorithm to accurately identify and extract valid structured data from the model output.

[0108] Step 4: semantic similarity calculation, implement semantic similarity calculation based on vector embedding, which is used to match the structured information of plant and equipment names in the text to be extracted with the standardized names in the plant mapping table and equipment mapping table after disassembly from different system sources.

[0109] Step 401: define the vector representation of the generated text, specifically:

[0110] For a given text t, its vector is represented as:

[0111] V(t)=f(t)

[0112] Wherein, t is the structured information of the plant and equipment names in the text to be processed or the standardized names in the plant mapping table and equipment mapping table after disassembly from different system sources, and f is the encoding function, for example: model.encode(string).

[0113] Step 402: Design global and local similarity matching algorithms to support searching for the most similar items in different ranges, specifically:

[0114] a) For the query text q and the text set T = {t1, t2, ..., tn}, the similarity of each item is calculated according to the following formula:

[0115]

[0116] Among them, the query text q is the structured information of the plant and equipment names in the text to be processed, and the text set T is a large set of standardized names in the plant mapping table and equipment mapping table after disassembly from different system sources, where · represents the dot product, and |V| represents the Euclidean norm of the vector.

[0117] b) In an example, the most similar item in set T can be selected as the matching result according to the following formula, and then jump to step 6.

[0118] most_similar(q,T)=argmax t∈T (similarity(q,t))

[0119] In another example, the similarity is sorted in descending order, and the first M names are taken as the candidate name set c = {c1, c2, ..., cn}, and the subsequent steps 403 to 504 are performed.

[0120] Step 403: Implement optimization strategies for vector computing to improve computing efficiency during large-scale data processing. The optimization strategies used include: using efficient vector library functions; using parallel and distributed computing to improve computing efficiency; optimizing memory resource allocation, etc.

[0121] Step 5: Standardize plant and equipment names. Match the target fields of the extracted equipment and plant names with the standardized mapping table to achieve data standardization.

[0122] Standardized algorithm for plant and station names, handling various naming variants and abbreviations. Includes:

[0123] a) Edit distance calculation: For two strings s1 and s2, their edit distance D(i, j) is calculated by the following method:

[0124]

[0125] Wherein, the string s1 is the structured information of the plant and equipment names in the text to be processed, s2 is any one of the candidate names c = {c1, c2, ..., cn}; D(i, j) is the semantic similarity distance;

[0126] b) Standardized score calculation: For candidate name c and standard name s, the standardized score that comprehensively considers semantic similarity and edit distance can be expressed as:

[0127] score(c,s)=α·similarity(c,s)+(1-α)·(1-normalized_edit_distance(c,s))

[0128] Among them, α is the weight parameter, similarity(c,s) is the result of calculating c and s according to the similarity formula, and normalized_edit_distance is the normalized edit distance D(i, j).

[0129] c) Best match selection: For a candidate name c and a standard name set s, the best match selection can be expressed as:

[0130] best_match(c, S) = argmax s∈S score(c,s)

[0131] At this point, the result of the second best matching method is obtained, and a unique mapping index table between the target fields of all plant and equipment names and the standard names is established.

[0132] Step 502: Call the algorithm in step 501 to process the standardized matching of device names, which can be flexibly processed according to different scenarios.

[0133] Step 503: Design a context-based mapping table filtering mechanism to narrow the equipment matching range according to the plant station matching information and improve the matching efficiency and accuracy.

[0134] Step 504: For some special cases, the calculation weights of the edit distance and the normalized score are adjusted to adapt to the special naming case.

[0135] Step 6: Data integration and output: integrate the standardized structured data and output it in the specified format.

[0136] Step 601: Design a data integration processing algorithm to integrate the information extracted from the large model and the standardized data.

[0137] Step 602: Implement a general JSON serialization method to ensure that all data types (including special numeric types) can be correctly converted into the JSON format.

[0138] Step 603: Develop an incremental data processing and output mechanism to support efficient processing and real-time output of large-scale data.

[0139] Step 604: Implement the data upload function and securely transmit the processing results back to the remote server for use by the maintenance system.

[0140] {"id":"XXXX-XXXX","station":{"name":"XX.XX Factory","processed_name":"XXXX Factory","confirm_name":"XXXX Factory","confirm_id":"9XXXXXX","similarity":0.8283858299255371},"devices":[{"name":"Unit X","processed_name":"Unit X","confirm_name":"Unit X of XXXX Factory","confirm_id":"9XXXXX6","similarity":0.5086394548416138},{"name":"Unit X","processed_name":"Unit X","confirm_name":"Unit X","confirm_id":"9XXXXXXX6","similarity":0.6201760768890381}]}.

[0141] Key technical points:

[0142] 1. Large model application: Using large language models to extract information from unstructured text improves the ability to understand complex expressions.

[0143] 2. Design dedicated instruction templates to guide the large model to accurately understand and extract key information from power maintenance data.

[0144] 3. Semantic similarity calculation: Use the pre-trained text embedding model to calculate semantic similarity and achieve high-precision text matching.

[0145] 4. Multi-level mapping mechanism: Through the multi-level mapping structure, the flexibility and accuracy of the standardization process are improved. In set theory, M = {M_original, M_processed, M_station, M_device}, where each Mi is a mapping function, Mi: X→Y

[0146] 5. Context-aware matching: Consider the site information during device name matching to improve matching accuracy.

[0147] 6. Special scenario processing: Implement customized processing logic for specific scenarios. It can be expressed as conditional probability: P(correct_match|special_scenario)>P(correct_match|general_scenario)

[0148] 7. Incremental processing: Supports efficient processing of large-scale data through item-by-item processing and instant output.

[0149] It should be noted that the present invention is not limited to the embodiments described above. The above specific embodiments are merely illustrative and not restrictive. Without departing from the scope of the present invention and the scope of protection of the claims, a person skilled in the art may make many forms of specific changes under the guidance of the present invention, which are all within the scope of protection of the present invention.

Claims

1. A structured extraction system for power maintenance data based on a large model, characterized in that: The following system is built to accurately match the extracted information with standardized equipment and plant names through semantic understanding and multi-level matching mechanism for complex heterogeneous texts from multiple sources, and provide them to maintenance users. The system includes: Data acquisition preprocessing module, collecting raw power maintenance data, data cleaning and formatting; A standardized mapping table preparation module is used to load and parse the standardized naming mapping table of power system equipment; The unstructured text feature extraction module configures and calls the large language model to extract features from the preprocessed text to be analyzed and obtain the structured information of the target field; The semantic similarity module uses semantic similarity calculation based on vector embedding to match the structured information of the target field and search the standardized naming mapping table to obtain the corresponding preliminary matching results; The plant and equipment name standardization module further calculates the preliminary matching results according to the standardization algorithm, establishes the standardized mapping index table of the current plant and each level of equipment in the plant, and realizes data standardization; The data integration and output module integrates the standardized mapping index table, encapsulates it into a system service or offline file, and outputs it to the power grid business system for use.

2. According to claim 1, a structured extraction method for power maintenance data based on a large model is characterized in that: The power maintenance data is multi-source heterogeneous text records, Excel tables, and database data; The data acquisition preprocessing module includes: The secure transmission unit uses a secure file transfer protocol to obtain data files from a remote server; it is used to prevent eavesdropping, tampering or forgery during the transmission process; A local cache unit, used for local cache of data files; The data parsing unit parses the multi-source heterogeneous data into a unified data format; including: using regular expressions to filter out a number of target data that meet the conditions in the data set, splitting or searching the target data according to the specified characters, and obtaining a set of segmented target data; The data cleaning unit is used to implement data cleaning, including removing irrelevant characters, unifying separators, and format standardization.

3. According to claim 1, a structured extraction system for power maintenance data based on a large model is characterized in that: The standardized mapping table preparation module includes: The multi-level mapping structure unit is used to collect the original mapping table, and parse and process the ID mapping data of the corresponding level into: plant station mapping table and equipment mapping table; it is used to meet the needs of different scenarios; wherein the original mapping table is a standardized naming mapping table of existing power system equipment; wherein each standardized naming mapping table stores the standard naming names of all plants and equipment under the current system key value without hierarchy; the parsing operation is to parse each entry of the original mapping table of the current plant station into the plant station mapping table and the equipment mapping table according to the hierarchy; A mapping data loading unit, used for reading ID mapping data from a text file; A rule unit is used to define the consistency of the naming rules of each standardized name in the ID mapping data and the data to be analyzed; and to match the mapping key value of each standardized name in the ID mapping data with the data to be analyzed; wherein the mapping key value is a classification identifier of data from different sources, and the consistent naming rules are used to match data from different sources.

4. The structured extraction system of power maintenance data based on a large model according to claim 1 is characterized in that: The unstructured text feature extraction module includes: The large language model unit is used to adopt the large language model and combine it with the parameter unit to iteratively train the input data to obtain an ideal model, so as to extract the structured information of the target field; the key target fields include: XX transformer, XX transformer substation, XX plant, XX power plant, XX wind farm, and XX wind farm; The parameter unit is used to define the roles, tasks, instruction templates, output data structures, and domain knowledge of the large model; the tasks, instruction templates, output data structures, and domain knowledge under different roles are different; the domain knowledge inputs common equipment types and maintenance terms in the power system into the model to help the model learn domain-specific expressions.

5. The structured extraction system of power maintenance data based on a large model according to claim 1 is characterized in that: The semantic similarity module includes: The semantic similarity calculation unit implements semantic similarity calculation based on vector embedding, which is used for matching and searching the structured information of the plant and equipment names in the text to be extracted with the standardized names in the plant and equipment mapping tables parsed from different system sources. A text vector representation unit is used to represent a given text t vector as: V(t)=f(t); wherein t is the structured information of the plant station and equipment name in the text to be processed or the standardized name in the plant station mapping table and equipment mapping table after parsing under different system sources, and f is an encoding function; The similarity calculation unit is used to calculate the similarity and search and match the most similar items in different ranges.

6. A structured extraction system for power maintenance data based on a large model according to claim 5, characterized in that: The similarity calculation includes: for the query text q and the text set T = {t1, t2, ..., tn}, calculating the similarity of each item according to the following formula: Among them, the query text q is the structured information of the plant and equipment names in the text to be processed, and the text set T is a large set of standardized names in the plant mapping table and equipment mapping table parsed from different system sources, where · represents the dot product, and |V| represents the Euclidean norm of the vector.

7. A structured extraction system for power maintenance data based on a large model according to claim 5, It is characterized in that The searching and matching of the most similar items in different ranges includes: Select the most similar item in set T as the matching result according to the following formula; Calculate most_similar(q, T) = argmax t∈T (similarity(q,t)); Alternatively, the similarity values ​​are arranged in descending order, and the first M names are taken as the candidate name set c = {c1, c2, ..., cn} as the matching result output.

8. A structured extraction system for power maintenance data based on a large model according to claim 5, It is characterized in that The standardization algorithm of the plant and equipment name standardization module is used to handle various naming variants and abbreviations; including: a) Edit distance calculation: For two strings s1 and s2, calculate the edit distance D(i, j): Wherein, the string s1 is the structured information of the plant and equipment names in the text to be processed, s2 is any one of the candidate names c = {c1, c2, ..., cn}; D(i, j) is the semantic similarity distance; b) For candidate name c and standard name s, semantic similarity and edit distance are comprehensively considered to calculate the standardized score: score(c,s)=α·similarity(c,s)+(1-α)·(1-morphalized_edit_distance(c,s)) Among them, α is the weight parameter, similarity(c,s) is the result of calculating c and s according to the similarity formula, and normalized_edit_distance is the normalized edit distance D(i, j); c) For a candidate name c and a standard name set s, the selection of the best match is expressed as: best_match(c,S)=argmax s∈S score(c,s) At this point, a unique standardized mapping index table of the target fields and standard names of all plant and equipment names is established.

9. The structured extraction system of power maintenance data based on a large model according to claim 1 is characterized in that: The data integration and output module uses the JSON serialization method to format the data in the standardized mapping index table; uses an incremental data processing and output mechanism to support efficient processing and real-time output of large-scale data; is used to implement the data upload function and securely transmit the processing results back to the remote server for use by the maintenance system.

10. A structured extraction method for power maintenance data based on a large model, characterized in that: The following steps are performed to identify the multi-source complex and heterogeneous power maintenance data, and a matching search is performed in the standardized mapping table to obtain a unique standard name and provide it to the maintenance user. The method includes the following steps: Step S1, collecting original power maintenance data, cleaning and formatting the data; Step S2, loading and parsing the standardized naming mapping table of power system equipment; Step S3: configure and call a large language model to extract features from the preprocessed text to be analyzed to obtain structured information of the target field; Step S4: using semantic similarity calculation based on vector embedding to match the structured information of the target field and search the standardized naming mapping table to obtain the corresponding preliminary matching result; Step S5: further calculate the preliminary matching results according to the standardized algorithm, establish a standardized mapping index table for the current plant and each level of equipment in the plant, and realize data standardization; Step S6: Integrate the standardized mapping index table, encapsulate it into a system service or an offline file, and output it to the power grid business system for use.

Images