A large model-based power data dynamic verification method

Abstract

The present application relates to the technical field of data processing, and particularly relates to a power data dynamic checking method based on a large model. Power multi-modal data to be checked is acquired, the power multi-modal data is fused and understood through a pre-trained multi-modal large model to acquire multi-dimensional unified semantic representation; a double-modal rule knowledge base containing a prior rule mode and a historical case mode is constructed, and a set of applicable checking rules is matched from the double-modal rule knowledge base based on the multi-dimensional unified semantic representation; the multi-dimensional unified semantic representation is subjected to first saliency checking based on the set of applicable checking rules, explicit data defects are identified and marked, and a first saliency checking result is generated; second saliency checking is performed based on the first saliency checking result and a pre-constructed multi-modal correlation graph to acquire a second saliency checking result; and the present application can greatly improve the reliability of power data quality and data application.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, and in particular to a dynamic verification method for power data based on a large model. Background Technology

[0002] Traditional power data verification methods primarily rely on manual verification and single-modal rule-based verification. These methods are suitable for verifying the compliance and accuracy of small amounts of single-modal power data, depending on fixed rules or human experience. However, traditional methods have significant limitations. They are difficult to adapt to cross-modal unified verification of multimodal data, lack historical case support leading to incomplete coverage and the potential to miss hidden risks. Furthermore, they are inefficient, lack intelligence and flexibility, and produce fragmented output results, making it difficult to meet the needs of efficient and accurate verification of massive amounts of power data. Summary of the Invention

[0003] This invention addresses the technical problems of existing technologies, such as incomplete coverage, easy omission of hidden risks, low verification efficiency, insufficient intelligence and flexibility, and scattered output results, by providing a dynamic verification method for power data based on a large model.

[0004] The technical solution of the present invention to solve the above-mentioned technical problems is as follows:

[0005] In a first aspect, the present invention provides a dynamic verification method for power data based on a large model, comprising: acquiring power multimodal data to be verified; performing fusion understanding on the power multimodal data through a pre-trained multimodal large model to obtain a multidimensional unified semantic representation, wherein the multidimensional unified semantic representation is associated with a corresponding basic data summary; constructing a bimodal rule knowledge base including prior rule modalities and historical case modalities; and matching an applicable verification rule set from the bimodal rule knowledge base based on the multidimensional unified semantic representation.

[0006] Based on the applicable verification rule set, a first saliency verification is performed on the multidimensional unified semantic representation to identify and mark explicit data defects, and generate a first saliency verification result.

[0007] Based on the first saliency check result and the pre-constructed multimodal association map, a second saliency check is performed to obtain the second saliency check result; the first saliency check result and the second saliency check result are merged to generate a structured check report.

[0008] Optionally, the power multimodal data includes at least two types of data: text data, tabular data, and image data. Specifically, the text data is converted into a standard structure of data name-value-unit-paragraph position; the tabular data is converted into a standard structure of data name-row identifier-column identifier-value-unit; and the image data is converted into a standard structure of data name-image identifier-data series-value-unit-trend description.

[0009] Optionally, the multimodal large model includes:

[0010] The input layer includes a text input submodule, a table input submodule, and an image input submodule.

[0011] A cross-modal coding layer is used to receive and encode the power multimodal data from the input layer. The cross-modal coding layer includes a text Transformer encoder, a table Transformer encoder, and a graph Transformer encoder.

[0012] A multimodal fusion layer is used to map the encoding output of the cross-modal coding layer to a high-dimensional semantic space of a preset dimension to generate the multi-dimensional unified semantic representation.

[0013] The output layer is used to extract the basic data summary according to the multidimensional unified semantic representation.

[0014] Optionally, a bimodal rule knowledge base is constructed, comprising a prior rule modality and a historical case modality. This includes: defining collection channels and scope, collecting corresponding prior rule information, and performing data unification processing on the prior rule information to obtain original prior rules; atomically splitting and standardizing the original prior rules based on preset semantic tags to obtain atomic prior rules; performing semantic embedding processing on the atomic prior rules to obtain the prior rule modality; obtaining historical verification cases and extracting multiple typical cases based on preset frequency constraints; and performing semantic embedding processing on the multiple typical cases to obtain the historical case modality.

[0015] Optionally, matching an applicable verification rule set from the bimodal rule knowledge base based on the multidimensional unified semantic representation includes: establishing a keyword index based on the bimodal rule knowledge base; performing preliminary matching based on the keyword index and the basic data summary of the multidimensional unified semantic representation to obtain a rule candidate set, wherein the rule candidate set is associated with a preliminary matching degree; vectorizing the rule candidate set and calculating the similarity with the multidimensional unified semantic representation; weightedly fusing the preliminary matching degree and the similarity, and determining the applicable verification rule set based on a preset selection threshold.

[0016] Optionally, performing a first saliency check on the multidimensional unified semantic representation based on the applicable verification rule set, identifying and marking explicit data defects, and generating a first saliency check result includes: establishing a mapping relationship between the applicable verification rule set and the multidimensional unified semantic representation;

[0017] Based on the mapping relationship, the applicable verification rule set and the multidimensional unified semantic representation are traversed to perform field-level information alignment; based on the alignment result, multidimensional quantization residuals are calculated, and cross-modal fusion is performed on the multidimensional quantization residuals to obtain the fusion difference quantization value; the fusion difference quantization value is compared with the preset difference significance constraint value to obtain the first significance verification result.

[0018] Optionally, based on the first saliency verification result and the pre-constructed multimodal association graph, a second saliency verification is performed to obtain the second saliency verification result, including: performing association verification on the semantic representations of the data with explicit defects marked in the first saliency verification result in combination with the multimodal association graph to generate a first association verification result; performing temporal consistency analysis on the semantic representations of the data without explicit defects marked in the first saliency verification result, and performing adaptive verification on the data to be evaluated without explicit defects marked in the first saliency verification result based on the temporal consistency analysis result to obtain a second association verification result; and outputting the first association verification result and the second association verification result as the second saliency verification result.

[0019] Specifically, the semantic representations of the data with explicit defects marked in the first saliency verification result are combined with the multimodal association graph to perform association verification and generate a first association verification result, including: extracting the semantic representations of the data with explicit defects marked in the first saliency verification result as a first type of representation.

[0020] Based on the multimodal association graph, the associated modal data of the first type of representation are determined; according to the bimodal rule knowledge base, the association verification rule set is determined, and the saliency verification of the associated modal data is performed to generate the first association verification result.

[0021] The process involves: performing temporal consistency analysis on the semantic representations of unmarked explicit data defects in the first saliency verification result; and, based on the temporal consistency analysis results, performing adaptive verification on the data to be evaluated that lacks explicit data defects in the first saliency verification result to obtain a second association verification result. This includes: extracting the semantic representations of unmarked explicit data defects in the first saliency verification result as a second type of representation, and performing implicit risk assessment based on the predicted trend boundary to obtain an implicit risk assessment value; if the implicit risk assessment value is higher than a first confidence threshold, determining the associated modal data of the second type of representation based on the multimodal association graph and performing saliency verification, wherein the modal depth of the associated modal data is positively correlated with the implicit risk assessment value, and obtaining the second association verification result; if the implicit risk assessment value is lower than the first confidence threshold but higher than the second confidence threshold, performing probabilistic saliency verification on the second type of representation based on the probability calculated from the historical error probability of the second type of representation and the preset business importance weight, and obtaining the second association verification result.

[0022] Optionally, the basic data summary includes at least the structured data name, modal source, original source, original value, and related data relationship.

[0023] By implementing this invention, it is possible to acquire power multimodal data to be verified, and to fuse and understand the power multimodal data through a pre-trained multimodal large model to obtain a multidimensional unified semantic representation. The multidimensional unified semantic representation is associated with a corresponding basic data summary, breaking down the format barriers of different modal data, realizing a unified understanding and representation of the data, and the basic data summary provides accurate data support for subsequent rule matching, thereby improving the targeting of verification.

[0024] By implementing this invention, it is possible to construct a bimodal rule knowledge base that includes prior rule modalities and historical case modalities, and to match applicable verification rule sets from the bimodal rule knowledge base based on the multidimensional unified semantic representation. By combining established rules with past experience, the coverage of verification rules can be broadened. Keyword indexing and similarity calculation ensure the accuracy and efficiency of rule matching and avoid invalid verification.

[0025] By implementing this invention, it is possible to perform a first saliency check on the multidimensional unified semantic representation based on the applicable verification rule set, identify and mark explicit data defects, generate a first saliency check result, accurately locate intuitively visible data errors, narrow the scope of subsequent verification, focus on key points, improve verification efficiency, and reduce redundant operations.

[0026] By implementing this invention, a second saliency check can be performed based on the first saliency check result and the pre-constructed multimodal association map to obtain the second saliency check result. This overcomes the limitations of a single explicit check, takes into account both the influence of association data and implicit risks, and dynamically adjusts the check strategy according to the risk level, thereby improving the comprehensiveness and depth of the check.

[0027] By implementing this invention, the first saliency verification result and the second saliency verification result can be merged to generate a structured verification report, making the verification results clear, easy to understand, and easy to trace, providing a clear basis for subsequent data correction and business optimization, and reducing the cost of interpreting the results.

[0028] In summary, by implementing this invention, comprehensiveness, accuracy, efficiency, and intelligence of power multimodal data verification can be achieved. It can accurately identify explicit data defects, deeply explore hidden risks, and output clear and actionable verification results, thereby significantly improving the quality of power data and the reliability of data applications. Attached Figure Description

[0029] Figure 1 A flowchart illustrating a dynamic verification method for power data based on a large model provided by the present invention;

[0030] Figure 2 The present invention provides a method for dynamic verification of power data based on a large model, which includes a flowchart of constructing a dual-modal rule knowledge base containing prior rule modalities and historical case modalities. Detailed Implementation

[0031] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0032] In the description of this invention, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more features. In the description of this invention, "a plurality of" means two or more, unless otherwise explicitly specified.

[0033] In the description of this invention, the term "for example" is used to mean "used as an example, illustration, or description." Any embodiment described as "for example" in this invention is not necessarily to be construed as being more preferred or advantageous than other embodiments. The following description is provided to enable any person skilled in the art to make and use the invention. Details are set forth in the following description for purposes of explanation. It should be understood that those skilled in the art will recognize that the invention can be made without using these specific details. In other instances, well-known structures and processes will not be described in detail to avoid obscuring the description of the invention with unnecessary detail. Therefore, the invention is not intended to be limited to the embodiments shown, but is consistent with the broadest scope of the principles and features disclosed herein.

[0034] Example 1, as Figure 1 As shown, this embodiment of the invention provides a dynamic verification method for power data based on a large model, including:

[0035] S100: Obtain the power multimodal data to be verified, and perform fusion understanding on the power multimodal data through a pre-trained multimodal large model to obtain a multidimensional unified semantic representation, wherein the multidimensional unified semantic representation is associated with a corresponding basic data summary;

[0036] S200: Construct a bimodal rule knowledge base containing prior rule modalities and historical case modalities, and match applicable verification rule sets from the bimodal rule knowledge base based on the multidimensional unified semantic representation;

[0037] S300: Perform a first saliency check on the multidimensional unified semantic representation based on the applicable check rule set, identify and mark explicit data defects, and generate a first saliency check result;

[0038] S400: Based on the first saliency verification result and the pre-constructed multimodal association map, perform a second saliency verification to obtain the second saliency verification result;

[0039] S500: Merge the first saliency check result and the second saliency check result to generate a structured check report.

[0040] S100: Obtain the power multimodal data to be verified, and perform fusion understanding on the power multimodal data through a pre-trained multimodal large model to obtain a multidimensional unified semantic representation. The multidimensional unified semantic representation is associated with a corresponding basic data summary, which includes at least the structured data name, modality source, original source, original value and related data relationship.

[0041] In this embodiment of the application, the purpose of step S100 is to break down the semantic fragmentation of multimodal data such as power text, tables, and images, generate a unified high-dimensional semantic representation, and at the same time extract structured basic data summaries to provide a unified and accurate data foundation for subsequent rule matching and difference recognition.

[0042] To achieve the above objectives, it is first necessary to obtain the power multimodal data to be verified.

[0043] In step S100 of this application embodiment, the power multimodal data includes at least two of the following: text data, tabular data, and image data:

[0044] The text data is converted into a standard structure of data name-value-unit-paragraph position;

[0045] The tabular data is converted into a standard structure of data name-row identifier-column identifier-value-unit;

[0046] The image data is converted into a standard structure of data name-image identifier-data series-value-unit-trend description.

[0047] Specifically, for text-based data, it is necessary to extract key power data from the text and organize it into a standard structure according to the order of data name, value, unit, and paragraph position.

[0048] For example, data can be extracted from the power project acceptance contract text. The data name is transformer rated capacity, the value is 100MVA, the unit is megavolt-ampere, and the paragraph position is the third paragraph of the first section of the fourth chapter.

[0049] For tabular data, it is necessary to parse electricity-related tabular data, clarify the data name, determine the corresponding row and column identifiers, extract the values ​​and units, and combine them into a standard structure.

[0050] For example, extract data from the power equipment parameter table. The data name is bus voltage, the row is labeled as row 5, the column is labeled as column 2, the value is 220kV, and the unit is kilovolt.

[0051] For image-based data, it is necessary to identify the core information of the power data in the image, determine the data name, image identifier, and data series to which it belongs, extract the numerical values ​​and units, describe the trends, and form a standard structure.

[0052] For example, data is extracted from the power load curve. The data is named "Daily Maximum Load", the image is labeled "Figure 3-2", the data series is "Summer Load Series", the value is 580MW, the unit is megawatt, and the trend is described as the load being higher from 8 am to 10 pm than at other times.

[0053] In step S100 of this application embodiment, the multimodal large model includes:

[0054] The input layer includes a text input submodule, a table input submodule, and an image input submodule.

[0055] A cross-modal coding layer is used to receive and encode the power multimodal data from the input layer. The cross-modal coding layer includes a text Transformer encoder, a table Transformer encoder, and a graph Transformer encoder.

[0056] A multimodal fusion layer is used to map the encoding output of the cross-modal coding layer to a high-dimensional semantic space of a preset dimension to generate the multi-dimensional unified semantic representation.

[0057] The output layer is used to extract the basic data summary according to the multidimensional unified semantic representation.

[0058] In this embodiment of the application, the purpose of obtaining the multimodal large model is to process different types of power multimodal data in a targeted manner through the hierarchical architecture of the multimodal large model, realize the encoding and deep fusion of cross-modal data, generate a unified multidimensional semantic representation, and extract structured basic data summaries to provide standardized and high-value data support for subsequent rule matching and data verification.

[0059] To achieve the above objectives, an input layer needs to be built first, which includes a text input submodule, a table input submodule, and an image input submodule.

[0060] The text input submodule receives standard text-based data, the table input submodule receives standard table-based data, and the image input submodule receives standard image-based data.

[0061] For example, the text input submodule receives data with the name "Line Loss" and the value "12kW" (unit: kilowatt), and the paragraph position "Chapter 2, Paragraph 3"; the table input submodule receives data with the name "Transformer Efficiency" and the values ​​"98.5%" (unit: percentage), with the row identifier "4" and column identifier "5".

[0062] Then, a cross-modal coding layer needs to be built. The cross-modal coding layer is used to receive the power multimodal data from the input layer and encode it. The cross-modal coding layer includes a text Transformer encoder, a table Transformer encoder, and a graph Transformer encoder.

[0063] The text Transformer encoder processes the data from the text input submodule, the table Transformer encoder processes the data from the table input submodule, and the graph Transformer encoder processes the data from the image input submodule, ultimately generating their respective modality coding vectors.

[0064] For example, the text Transformer encoder encodes line loss-related text data into a text encoding vector h_T, and the table Transformer encoder encodes transformer efficiency-related table data into a table encoding vector h_Tab.

[0065] Next, a multimodal fusion layer needs to be built. The multimodal fusion layer is used to map the encoding output of the cross-modal coding layer to a high-dimensional semantic space of a preset dimension, thereby generating the multi-dimensional unified semantic representation.

[0066] For example, if the high-dimensional semantic space is preset to 768 dimensions, the fusion layer maps h_T, h_Tab, and the image encoding vector h_G to this space to generate the fusion representation h_fusion.

[0067] Finally, an output layer needs to be built, which is used to extract the basic data summary according to the multidimensional unified semantic representation.

[0068] The basic data summary includes at least the structured data name, modal source, original source, original value, and related data relationship.

[0069] For example, extract the basic data summary from h_fusion, with the data name being line loss; the modal source being text; the original source being a power line operation and maintenance report; the original value being 12kW; and the correlation data relationship being positively correlated with the line length.

[0070] like Figure 2 As shown, in step S200 of this embodiment, a dual-modal rule knowledge base comprising a priori rule modality and a historical case modality is constructed, including:

[0071] Define the collection channels and collection scope, corresponding to the collection of prior rule information, and perform data unification processing on the prior rule information to obtain the original prior rules;

[0072] Based on preset semantic tags, the original prior rules are atomically decomposed and standardized to obtain atomic prior rules;

[0073] Perform semantic embedding processing on the atomized prior rules to obtain the prior rule modality;

[0074] Obtain historical verification cases and extract multiple typical cases based on preset frequency constraints;

[0075] Semantic embedding processing is performed on multiple typical cases to obtain the historical case modality.

[0076] In step S100 of this application embodiment, the construction of a dual-modal rule knowledge base provides comprehensive and accurate rule support for power data verification, solving the problems of traditional rule bases being single and having low matching accuracy. At the same time, through standardization and semantic embedding processing, the rules and cases have efficient retrieval and matching capabilities.

[0077] To achieve the above objectives, it is first necessary to define the collection channels and scope, collect corresponding prior rule information, and perform data unification processing on the prior rule information to obtain the original prior rules.

[0078] The defined channels for collecting prior rule information can include legitimate channels such as the official websites of relevant committees, the publishing platforms of industry regulatory authorities, and enterprise internal technical file management systems. The scope of collection includes power industry regulations, mandatory standards, recommended standards, equipment operation specifications, and data reporting specifications. The collected prior rule information undergoes unified data processing, removing headers, footers, watermarks, and non-regular explanatory text, converting it to standard UTF-8 encoded text to obtain the original prior rules.

[0079] For example, power data quality standards are collected from the official website of a certain bureau, and after unified processing, the original prior rule is obtained: the temperature rise of the transformer must not exceed 65K.

[0080] Next, based on preset semantic tags, the original prior rules need to be atomically split and standardized to obtain atomic prior rules.

[0081] This involves setting preset semantic tags, such as numerical, formatted, named, and logical types. Based on these semantic tags, the original prior rules are atomically decomposed into independent and complete rule units. These units are then standardized to clarify information such as rule identifier, type, text, and applicable domain, thus obtaining atomic prior rules.

[0082] For example, the original prior rule "110kV line insulation resistance value should be ≥1000MΩ" is split into numerical semantic tags and standardized to obtain an atomized prior rule: Rule identifier: numerical-2024-003; Rule type: numerical; Rule text: 110kV line insulation resistance value is not less than 1000MΩ; Applicable field: power line inspection; Verification parameter: lower threshold 1000MΩ; Tolerance accuracy: 0.1MΩ.

[0083] Next, semantic embedding processing needs to be performed on the atomic prior rules to obtain the prior rule modality. That is, semantic embedding processing is performed on the atomic prior rules to convert the standardized rule text into a fixed-length feature vector, with the vector dimension consistent with the dimension of the multimodal unified semantic representation. For example, the atomic prior rules related to the insulation resistance of the 110kV line mentioned above are embedded to generate a 768-dimensional feature vector, which serves as the prior rule modality.

[0084] Furthermore, historical verification cases are obtained, and multiple typical cases are extracted in combination with preset frequency constraints.

[0085] This involves collecting historical verification cases of power data, covering verification scenarios for multimodal data such as text, tables, and images. Then, preset frequency constraints are set, such as selecting similar verification cases that have appeared ≥5 times in the past 3 years, and extracting multiple typical cases accordingly.

[0086] For example, among the cases involving transformer load data verification in the past three years, those cases where the actual operating load of the transformer with a rated load of 31.5 MVA exceeded the rated value by 15% on eight occasions were identified as data anomalies and were used as typical cases.

[0087] Finally, semantic embedding processing is performed on multiple typical cases to obtain the historical case modality.

[0088] This involves semantic embedding of multiple extracted typical cases, converting the core information in the cases into feature vectors that are consistent with the prior rule modality dimension, and obtaining the historical case modality.

[0089] For example, the typical cases of transformer load anomalies mentioned above are embedded and processed to generate a 768-dimensional feature vector, which serves as a historical case mode.

[0090] In step S200 of this application embodiment, matching an applicable verification rule set from the bimodal rule knowledge base based on the multidimensional unified semantic representation includes:

[0091] Based on the aforementioned bimodal rule knowledge base, a keyword index is established;

[0092] Based on the keyword index, preliminary matching is performed using the basic data summary of the multidimensional unified semantic representation to obtain a rule candidate set, wherein the rule candidate set is associated with a preliminary matching degree;

[0093] Vectorize the candidate set of rules and calculate the similarity with the multidimensional unified semantic representation accordingly;

[0094] The preliminary matching degree and the similarity degree are weighted and fused together, and the applicable verification rule set is determined by combining them with a preset selection threshold.

[0095] In step S200 of this application embodiment, the purpose of the above steps is to accurately select verification rules that match the power multimodal data from the dual-modal rule knowledge base. By combining the initial screening with keyword indexing and the accurate calculation of vector similarity, the efficiency and accuracy of rule matching are improved, ensuring that subsequent verification has targeted rule support.

[0096] To achieve the above objectives, a keyword index must first be established based on the aforementioned bimodal rule knowledge base. This involves extracting core keywords from the prior rule modalities and historical case modalities of the bimodal rule knowledge base. These keywords include power industry terms, indicator names, verification objects, and key qualifiers. An inverted index is then created for each keyword, recording the rule identifiers and their locations containing that keyword.

[0097] For example, extract the keyword "transformer temperature rise 65K" from the rule "transformer temperature rise shall not exceed 65K", and create an index to associate it with the rule identifier: value-2024-005.

[0098] Next, based on the keyword index and the basic data summary of the multidimensional unified semantic representation, preliminary matching is performed to obtain a rule candidate set, wherein the rule candidate set is associated with a preliminary matching degree. This involves extracting core information from the basic data summary corresponding to the multidimensional unified semantic representation, such as data name, modality source, and association relationships. Combining the keyword index, the basic data summary is preliminarily matched with rules and cases in the knowledge base to filter out rules with high relevance, forming a rule candidate set, and the preliminary matching degree of each candidate rule is labeled.

[0099] For example, the basic data summary is: Data Name - Transformer Temperature Rise - Original Value - 72K. Through keyword indexing, three rules containing the keywords "transformer" and "temperature rise" are matched, with initial matching degrees of 0.82, 0.75 and 0.61 respectively, forming a rule candidate set.

[0100] Next, the candidate rule set needs to be vectorized, and the similarity with the multidimensional unified semantic representation needs to be calculated accordingly. That is, each rule in the candidate rule set is vectorized, converted into a feature vector with the same dimensions as the multidimensional unified semantic representation. Specifically, a cosine similarity algorithm can be used to calculate the similarity between the candidate rule vector and the multidimensional unified semantic representation vector.

[0101] For example, the vector corresponding to the rule candidate set that the transformer temperature rise must not exceed 65K is V1, and the multidimensional unified semantic representation vector is V2. The similarity is calculated to be 0.88.

[0102] Finally, the preliminary matching degree and the similarity degree need to be weighted and fused together, and the applicable verification rule set needs to be determined in combination with the preset selection threshold.

[0103] Specifically, a retrieval weight coefficient ω needs to be set to balance the proportion of initial matching degree and similarity, for example, ω=0.4. Then, the comprehensive matching score of each candidate rule is calculated according to the formula MatchScore=ω×initial matching degree+(1-ω)×similarity. A preset selection threshold is set, such as 0.7, and candidate rules with a comprehensive matching score ≥ the selection threshold are included in the applicable validation rule set.

[0104] For example, a candidate rule has an initial matching degree of 0.82, a similarity of 0.88, and ω=0.4. The calculated MatchScore is 0.4×0.82+0.6×0.88=0.856, which is higher than the selection threshold of 0.7, so it is included in the applicable validation rule set.

[0105] In step S300 of this application embodiment, a first saliency check is performed on the multidimensional unified semantic representation based on the applicable check rule set to identify and mark explicit data defects and generate a first saliency check result, including:

[0106] Establish a mapping relationship between the applicable verification rule set and the multidimensional unified semantic representation;

[0107] Based on the mapping relationship, the applicable verification rule set is traversed and field-level information is aligned with the multidimensional unified semantic representation.

[0108] Based on the alignment results, multi-dimensional quantization residuals are calculated, and cross-modal fusion is performed on the multi-dimensional quantization residuals to obtain the fusion difference quantization value.

[0109] The first significance verification result is obtained by comparing the fusion difference quantification value with the preset difference significance constraint value.

[0110] In this embodiment of the application, the purpose of step S300 is to identify and mark the obvious defects in the power data by accurately connecting the applicable verification rule set with the multi-dimensional unified semantic representation, so as to provide a clear problem orientation for subsequent secondary verification and ensure the accuracy and pertinence of the verification results.

[0111] To achieve the above objectives, it is first necessary to establish a mapping relationship between the applicable verification rule set and the multidimensional unified semantic representation. That is, to clarify the verification object and standard requirements of each rule in the applicable verification rule set, and to establish a one-to-one correspondence with the indicators and parameters in the multidimensional unified semantic representation.

[0112] For example, a mapping relationship is established between the applicable verification rule "110kV line insulation resistance value ≥ 1000MΩ" and the semantic representation "data name - 110kV - line insulation resistance".

[0113] Next, based on the mapping relationship, it is necessary to traverse the applicable verification rule set and the multidimensional unified semantic representation to perform field-level information alignment; that is, according to the mapping relationship, traverse each applicable rule and its corresponding semantic representation, and align the fields such as data name, value, unit, and logical relationship one by one.

[0114] For example, after unifying the units of "threshold lower limit 1000MΩ, unit MΩ" in the rule and "original value 950MΩ, unit megohm" in the semantic representation, the numerical field and the threshold field are aligned.

[0115] Then, based on the alignment results, multi-dimensional quantization residuals are calculated, and cross-modal fusion is performed on the multi-dimensional quantization residuals to obtain the fusion difference quantization value.

[0116] Based on the alignment results, multi-dimensional quantization residuals are calculated, such as numerical bias and format consistency bias. Cross-modal fusion is then performed on these multi-dimensional quantization residuals, integrating the differences between different dimensions through methods such as weighted summation to obtain the quantized value of the fused difference.

[0117] For example, if the numerical residual of a certain data is 0.05, the format residual is 0, and the logical residual is 0.02, and the data is fused with weights of 0.6, 0.2, and 0.2, the fused difference quantification value is 0.05×0.6+0×0.2+0.02×0.2=0.034.

[0118] Finally, the fused difference quantification value is compared with the preset difference significance constraint value to obtain the first significance verification result. A preset difference significance constraint value, such as 0.05, must first be set. If the fused difference quantification value is greater than the difference significance constraint value, an explicit data defect is determined and marked; if it is less than or equal to the difference significance constraint value, the data is determined to be normal.

[0119] For example, if the fusion difference quantification value is 0.034 ≤ the difference significance constraint value is 0.05, the data is determined to have no explicit defects; if the fusion difference quantification value is 0.06 > 0.05, it is marked as an explicit defect of numerical excess.

[0120] In step S400 of this application embodiment, based on the first saliency verification result and the pre-constructed multimodal association map, a second saliency verification is performed to obtain the second saliency verification result, including:

[0121] The semantic representations that mark explicit data defects in the first saliency verification result are combined with the multimodal association graph for association verification to generate the first association verification result.

[0122] The semantic representation of unmarked explicit data defects in the first saliency verification result is subjected to temporal consistency analysis. Based on the temporal consistency analysis result, the data to be evaluated that are unmarked explicit data defects in the first saliency verification result are subjected to adaptive verification to obtain the second association verification result.

[0123] The first correlation verification result and the second correlation verification result are output as the second saliency verification result.

[0124] In this embodiment of the application, the purpose of step S400 is to, based on the first saliency verification, uncover the correlation between explicit defects and implicit data anomalies, make up for the limitations of a single explicit verification, realize in-depth verification of power data, and improve the comprehensiveness and accuracy of the verification results.

[0125] Specifically, the semantic representations of the data with explicit defects marked in the first saliency verification result need to be combined with the multimodal association graph for association verification to generate the first association verification result.

[0126] In step S400 of this application embodiment, the semantic representations of the first saliency verification results that are marked with explicit data defects are combined with the multimodal association graph to perform association verification and generate a first association verification result, including:

[0127] Extract the semantic representations that mark the existence of explicit data defects in the first saliency verification results as the first type of representation;

[0128] The associated modal data of the first type of representation are determined based on the multimodal association map;

[0129] Based on the bimodal rule knowledge base, a set of association verification rules is determined, and the saliency of the association modal data is verified to generate a first association verification result.

[0130] In embodiment S400 of this application, the purpose of the above-mentioned detailed steps is to uncover the chain-related problems behind explicit data defects, avoid only treating a single defect in isolation and missing the implicit or explicit anomalies of related modal data, and improve the depth and completeness of power data verification.

[0131] To achieve this goal, firstly, it is necessary to extract the semantic representations of those marked as having explicit data defects from the first saliency check results as the first type of representation. That is, from the first saliency check results, all semantic representations marked as having explicit data defects are screened and extracted, and these are defined as the first type of representation.

[0132] For example, in the first saliency verification result, the label "Data Name - Main Transformer Oil Level, Value -35cm, lower than the rule threshold of 40cm" is an explicit defect, and this semantic representation is the first type of representation.

[0133] Next, it is necessary to determine the associated modal data of the first type of representation based on the multimodal association graph. That is, by calling the pre-constructed multimodal association graph and retrieving other modal data that have logical associations and data correspondences with the entity information of the first type of representation, such as device entities and indicator entities.

[0134] Specifically, the associated modal data can come from different modalities such as text, tables, and images, and cover related data such as associated indicators, associated devices, and associated time points.

[0135] For example, the first type of characterization is the abnormal oil level of the main transformer. By querying the correlation graph, the correlation modal data "main transformer oil temperature -85℃, cooling system operating status - fan not started, oil level sensor detection voltage -2.1V" were found, which came from the table, text and image modal respectively.

[0136] Then, based on the bimodal rule knowledge base, it is necessary to determine the association verification rule set, perform saliency verification on the association modal data, and generate the first association verification result.

[0137] This involves retrieving the rules corresponding to the associated modal data from the bimodal rule knowledge base, i.e., the association verification rule set, and performing verification according to the standard procedure for first significance verification. Specifically, this includes establishing the mapping relationship between rules and associated modal data, aligning field-level information, calculating the fusion difference quantification value, and comparing it with the difference significance constraint value.

[0138] For example, the associated modal data shows the transformer oil temperature as -85℃, and the corresponding verification rule is that the transformer's normal operating oil temperature is ≤80℃, with a significant difference constraint value of 0.05. The calculated fusion difference quantification value is 0.08, which is greater than the significant constraint value, and is therefore judged as an explicit defect. The cooling system operating status - fan not started - corresponds to the rule that the fan should start automatically when the oil temperature is ≥75℃, and is judged as a logical explicit defect. These results are integrated to generate the first associated verification result.

[0139] Furthermore, it is necessary to perform temporal consistency analysis on the semantic representation of the unmarked explicit data defects in the first saliency verification result, and based on the temporal consistency analysis result, to perform adaptive verification on the data to be evaluated that are unmarked explicit data defects in the first saliency verification result, and obtain the second association verification result.

[0140] In step S400 of this application embodiment, the semantic representation of the unmarked explicit data defects in the first saliency verification result is subjected to temporal consistency analysis, and based on the temporal consistency analysis result, adaptive verification is performed on the data to be evaluated that are unmarked explicit data defects in the first saliency verification result to obtain the second association verification result, including:

[0141] The semantic representation of the unmarked explicit data defects in the first significance verification result is extracted as the second type of representation, and the implicit risk assessment is carried out based on the predicted trend boundary to obtain the implicit risk assessment value.

[0142] If the latent risk assessment value is higher than the first confidence threshold, then the associated modal data of the second type of representation is determined according to the multimodal association map and significance verification is performed, wherein the modal depth of the associated modal data is positively correlated with the latent risk assessment value, and the second association verification result is obtained;

[0143] If the implicit risk assessment value is lower than the first confidence threshold but higher than the second confidence threshold, then the probability calculated based on the historical error probability of the second type of representation and the preset business importance weight is used as the verification probability. The second type of representation is then subjected to probabilistic significance verification to obtain the second association verification result.

[0144] In step S400 of this application embodiment, the purpose of the above-mentioned detailed steps is to identify hidden risks through time series consistency analysis for power data without obvious defects, and then perform differentiated adaptive verification according to the risk level to accurately capture cross-modal hidden anomalies and time series fluctuation problems, thereby further improving the comprehensiveness and depth of data verification.

[0145] To achieve the above objectives, it is first necessary to extract the semantic representation of the unmarked explicit data defects in the first significance verification result as the second type of representation, and then conduct implicit risk assessment based on the predicted trend boundary to obtain the implicit risk assessment value.

[0146] From the results of the first significance check, semantic representations of unlabeled explicit data defects are selected and defined as the second type of representation. Then, based on the preset predicted trend boundary, combined with historical data from the same period and industry standard trend range, the implicit risk assessment of the second type of representation is carried out, and the implicit risk assessment value in the range of 0-1 is output.

[0147] For example, the second type of characterization is a daily power supply of 520,000 kWh for a 10kV line, with a predicted trend boundary of 450,000-500,000 kWh. This data exceeds the boundary and there is no explicit rule conflict. The implicit risk assessment value is 0.87.

[0148] Next, if the latent risk assessment value is higher than the first confidence threshold, the associated modal data of the second type of representation is determined according to the multimodal association map and significance verification is performed, wherein the modal depth of the associated modal data is positively correlated with the latent risk assessment value, and the second association verification result is obtained.

[0149] Specifically, a first confidence threshold needs to be set, such as 0.85. If the latent risk assessment value is higher than this first confidence threshold, the associated modal data of the second type of representation is queried through the multimodal association graph. The modal depth of the associated modal data is positively correlated with the latent risk assessment value. Based on the applicable verification rule set, a significance check is performed on the associated modal data, and abnormal results are recorded. Here, "based on the applicable verification rule set" refers to using the set of verification rules previously matched and selected from the bimodal rule knowledge base that are highly correlated with the current power data to be verified as the core basis for performing the verification operation during the data verification process.

[0150] For example, the implicit risk assessment value is 0.87 > 0.85, corresponding to a modal depth of 3. The query found related data: line loss 8.2%, transformer load rate 92%, and user electricity consumption growth rate 15%. The verification found that the line loss exceeded the rule threshold of 7%, which was judged as an implicit association defect and included in the second association verification result.

[0151] Then, if the implicit risk assessment value is lower than the first confidence threshold but higher than the second confidence threshold, the probability calculated based on the historical error probability of the second type of representation and the preset business importance weight is used as the verification probability. The second type of representation is then subjected to probabilistic significance verification to obtain the second association verification result.

[0152] Specifically, a second confidence threshold needs to be set first, for example, 0.7. If the implicit risk assessment value is lower than the first confidence threshold but higher than the second confidence threshold, the verification probability is calculated. The verification probability = historical error probability × preset business importance weight. Then, a probabilistic significance check is performed on the second type of representation according to the verification probability. If the check is triggered, the complete check process is executed; if it is not triggered, the data is considered normal.

[0153] For example, the second type of characteristic has a main transformer power factor of 0.91, a latent risk assessment value of 0.78, a historical error probability of 8%, and a business importance weight of 0.7. Therefore, the verification probability is 8% × 0.7 = 5.6%. After triggering the verification, no abnormality is found, and the data is judged to be normal. The other characteristic has a verification probability of 20%. After triggering the verification, an abnormal logical association is found, and it is included in the second association verification result.

[0154] Finally, the first association verification result and the second association verification result are output as the second saliency verification result.

[0155] In step S500 of this application embodiment, it is also necessary to merge the first saliency verification result and the second saliency verification result to generate a structured verification report.

[0156] This involves summarizing explicit defects from the first significance test, related defects and latent anomalies from the second significance test, removing duplicate records, and labeling the error type and risk level of each result. The content is then organized logically according to the report cover, summary, difference details, error classification, and risk assessment, presenting core data using tables and charts. The final output is a standardized document including data location, rule basis, and revision suggestions, ensuring clear and traceable information.

[0157] It should be noted that the descriptions of each embodiment in the above embodiments have different focuses. For parts that are not described in detail in a certain embodiment, please refer to the relevant descriptions in other embodiments.

[0158] Those skilled in the art will understand that embodiments of the present invention can provide methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0159] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, as well as combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded computer, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0160] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0161] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0162] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Clearly, those skilled in the art can make various alterations and variations to the invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the invention and its equivalents, the invention is also intended to include these modifications and variations.

Claims

1. A dynamic verification method for power data based on a large model, characterized in that, include: The power multimodal data to be verified is acquired, and the power multimodal data is fused and understood through a pre-trained multimodal large model to obtain a multidimensional unified semantic representation, wherein the multidimensional unified semantic representation is associated with a corresponding basic data summary. Construct a bimodal rule knowledge base that includes prior rule modalities and historical case modalities, and match applicable verification rule sets from the bimodal rule knowledge base based on the multidimensional unified semantic representation; Based on the applicable verification rule set, a first saliency verification is performed on the multidimensional unified semantic representation to identify and mark explicit data defects, and generate a first saliency verification result. Based on the first saliency check result and the pre-constructed multimodal association map, a second saliency check is performed to obtain the second saliency check result; Merge the first saliency check result and the second saliency check result to generate a structured check report; The process of matching applicable verification rule sets from the bimodal rule knowledge base based on the multidimensional unified semantic representation includes: Based on the aforementioned bimodal rule knowledge base, a keyword index is established; Based on the keyword index, preliminary matching is performed using the basic data summary of the multidimensional unified semantic representation to obtain a rule candidate set, wherein the rule candidate set is associated with a preliminary matching degree; Vectorize the candidate set of rules and calculate the similarity with the multidimensional unified semantic representation accordingly; The preliminary matching degree and the similarity degree are weighted and fused together, and the applicable verification rule set is determined by combining them with a preset selection threshold.

2. The method for dynamic verification of power data based on a large model as described in claim 1, characterized in that, The power multimodal data includes at least two types: text data, tabular data, and image data, among which: The text data is converted into a standard structure of data name-value-unit-paragraph position; The tabular data is converted into a standard structure of data name-row identifier-column identifier-value-unit; The image data is converted into a standard structure of data name-image identifier-data series-value-unit-trend description.

3. The method for dynamic verification of power data based on a large model as described in claim 1, characterized in that, The multimodal large model includes: The input layer includes a text input submodule, a table input submodule, and an image input submodule. A cross-modal coding layer is used to receive and encode the power multimodal data from the input layer. The cross-modal coding layer includes a text Transformer encoder, a table Transformer encoder, and a graph Transformer encoder. A multimodal fusion layer is used to map the encoding output of the cross-modal coding layer to a high-dimensional semantic space of a preset dimension to generate the multi-dimensional unified semantic representation. The output layer is used to extract the basic data summary according to the multidimensional unified semantic representation.

4. The method for dynamic verification of power data based on a large model as described in claim 1, characterized in that, Construct a bimodal rule knowledge base that includes both prior rule modalities and historical case modalities, including: Define the collection channels and collection scope, corresponding to the collection of prior rule information, and perform data unification processing on the prior rule information to obtain the original prior rules; Based on preset semantic tags, the original prior rules are atomically decomposed and standardized to obtain atomic prior rules; Perform semantic embedding processing on the atomized prior rules to obtain the prior rule modality; Obtain historical verification cases and extract multiple typical cases based on preset frequency constraints; Semantic embedding processing is performed on multiple typical cases to obtain the historical case modality.

5. The method for dynamic verification of power data based on a large model as described in claim 1, characterized in that, Based on the applicable verification rule set, a first saliency check is performed on the multidimensional unified semantic representation to identify and mark explicit data defects, and a first saliency check result is generated, including: Establish a mapping relationship between the applicable verification rule set and the multidimensional unified semantic representation; Based on the mapping relationship, the applicable verification rule set is traversed and field-level information is aligned with the multidimensional unified semantic representation. Based on the alignment results, multi-dimensional quantization residuals are calculated, and cross-modal fusion is performed on the multi-dimensional quantization residuals to obtain the fusion difference quantization value. The first significance verification result is obtained by comparing the fusion difference quantification value with the preset difference significance constraint value.

6. The method for dynamic verification of power data based on a large model as described in claim 1, characterized in that, Based on the first saliency check result and the pre-constructed multimodal association graph, a second saliency check is performed to obtain the second saliency check result, including: The semantic representations that mark explicit data defects in the first saliency verification result are combined with the multimodal association graph for association verification to generate the first association verification result. The semantic representation of unmarked explicit data defects in the first saliency verification result is subjected to temporal consistency analysis. Based on the temporal consistency analysis result, the data to be evaluated that are unmarked explicit data defects in the first saliency verification result are subjected to adaptive verification to obtain the second association verification result. The first correlation verification result and the second correlation verification result are output as the second saliency verification result.

7. The method for dynamic verification of power data based on a large model as described in claim 6, characterized in that, The semantic representations of data with explicit defects in the first saliency verification result are combined with the multimodal association graph for association verification to generate a first association verification result, including: Extract the semantic representations that mark the existence of explicit data defects in the first saliency verification results as the first type of representation; The associated modal data of the first type of representation are determined based on the multimodal association map; Based on the bimodal rule knowledge base, a set of association verification rules is determined, and the saliency of the association modal data is verified to generate a first association verification result.

8. The method for dynamic verification of power data based on a large model as described in claim 6, characterized in that, For the semantic representation of unlabeled explicit data defects in the first saliency verification result, temporal consistency analysis is performed. Based on the temporal consistency analysis result, adaptive verification is performed on the data to be evaluated that are unlabeled explicit data defects in the first saliency verification result to obtain the second association verification result, including: The semantic representation of the unmarked explicit data defects in the first significance verification result is extracted as the second type of representation, and the implicit risk assessment is carried out based on the predicted trend boundary to obtain the implicit risk assessment value. If the latent risk assessment value is higher than the first confidence threshold, then the associated modal data of the second type of representation is determined according to the multimodal association map and significance verification is performed, wherein the modal depth of the associated modal data is positively correlated with the latent risk assessment value, and the second association verification result is obtained; If the implicit risk assessment value is lower than the first confidence threshold but higher than the second confidence threshold, then the probability calculated based on the historical error probability of the second type of representation and the preset business importance weight is used as the verification probability. The second type of representation is then subjected to probabilistic significance verification to obtain the second association verification result.

9. The method for dynamic verification of power data based on a large model as described in claim 1, characterized in that, The basic data summary includes at least the structured data name, modal source, original source, original value, and related data relationship.

Images