Power multi-modal knowledge base question-answer pair construction method and device

Abstract

The invention relates to a power multi-mode knowledge base question and answer pair construction method and device. The method comprises the following steps: screening at least two target electric power knowledge graph types, and constructing a plurality of business knowledge units based on target scene information of electric power production environment inspection business of each target electric power knowledge graph type; performing knowledge association analysis on the plurality of business knowledge units to obtain a logic mapping relationship among the plurality of business knowledge units; based on the logic mapping relationship, performing structured recombination on the plurality of business knowledge units to obtain a structured recombination result, and matching a corresponding question and answer sentence pattern template for the structured recombination result; and based on the question and answer sentence pattern templates adapted to the business knowledge units, converting the structured recombination result into candidate power knowledge question and answer pairs, verifying the candidate power knowledge question and answer pairs by using a large language model, and outputting a target power knowledge question and answer pair. By adopting the method, efficient answers can be generated for cross-type power inspection data.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of electric power operation and maintenance, and in particular to an electric power multi-modal knowledge base question and answer pair construction method and device, computer equipment, computer readable storage medium and computer program product. BACKGROUND

[0002] Electric power production environment inspection is a core link to ensure the safe and stable operation of the power system, and its efficiency and accuracy are directly related to the power grid operation and maintenance quality, fault response speed and power supply reliability. In the current electric power production environment inspection, a single modal analysis method based on a small model is often used, such as identifying equipment appearance defects only through an image small model or relying only on a text analysis tool to process maintenance records.

[0003] Although the traditional method has certain application value in a single data processing scene, it cannot realize effective association and integration of cross-type data due to the multi-modal characteristics of electric power inspection data (including images, videos, time series signals, text records, etc.), and highly depends on manual experience to generate maintenance schemes, which leads to false positives and false negatives in equipment defect identification due to environmental interference, one-sidedness in equipment state evaluation due to lack of multi-dimensional data support, low efficiency and poor consistency in maintenance scheme generation, and thus cannot meet the needs of intelligent and efficient production environment inspection, and restricts the overall efficiency improvement of electric power operation and maintenance.

[0004] Therefore, there is an urgent need for an electric power multi-modal knowledge base question and answer pair construction method, device, computer equipment, computer readable storage medium and computer program product, which can generate efficient answers for cross-type electric power inspection data. SUMMARY

[0005] Therefore, there is an urgent need for an electric power multi-modal knowledge base question and answer pair construction method, device, computer equipment, computer readable storage medium and computer program product, which can generate efficient answers for cross-type electric power inspection data.

[0006] In a first aspect, the present application provides an electric power multi-modal knowledge base question and answer pair construction method, comprising:

[0007] screening at least two target electric power knowledge graph types, and constructing a plurality of business knowledge units based on target scene information of electric power production environment inspection business of each target electric power knowledge graph type;

[0008] performing knowledge association analysis on the plurality of business knowledge units to obtain a logical mapping relationship between the plurality of business knowledge units;

[0009] Based on the logical mapping relationship, multiple business knowledge units are structurally reorganized to obtain the structural reorganization result, and the corresponding question-and-answer sentence template is adapted for the structural reorganization result;

[0010] Based on the question-and-answer sentence templates adapted to each business knowledge unit, the structured reorganization results are transformed into candidate power knowledge question-and-answer pairs, and the candidate power knowledge question-and-answer pairs are verified using a large language model to output the target power knowledge question-and-answer pairs.

[0011] In one embodiment, the step of performing knowledge association analysis on multiple business knowledge units to obtain the logical mapping relationship between the multiple business knowledge units includes:

[0012] For any two business knowledge units under the same target power knowledge graph type, based on the pre-built business association semantic dictionary, business semantic association matching is performed on the business feature items of the two business knowledge units to obtain the association matching result.

[0013] Based on the association matching results, two first business knowledge units that have business semantic association are selected as initial association candidate pairs. The number of business feature items that have business semantic association between the two first business knowledge units is divided by the total number of business feature items between the two first business knowledge units to obtain the feature association coverage rate.

[0014] Select initial candidate pairs of associations with a feature association coverage rate not less than a preset coverage rate threshold, and use them as target candidate pairs of associations;

[0015] Based on the business process of power production environment inspection, the business process dependency relationship between two business knowledge units in the target association candidate pair is analyzed. The business process dependency relationship includes pre-dependency relationship, post-dependency relationship and parallel relationship.

[0016] In one embodiment, the step of structurally reorganizing multiple business knowledge units based on the logical mapping relationship to obtain a structured reorganization result includes:

[0017] Determine the business process steps corresponding to all pre-dependencies and post-dependencies in the logical mapping relationship in the business process, and sort each business process step in the execution order to obtain a business process step sequence.

[0018] Based on the attribution relationship between multiple business knowledge units and each business process step in the business process step sequence, the attribution business knowledge unit group corresponding to each business process step is determined, and based on the parallel relationship, the parallel business knowledge unit group that is executed in parallel within each attribution business knowledge unit group is determined.

[0019] Based on the business process step sequence, the belonging business knowledge unit group, and the collaborative operation dimension of business knowledge units within the parallel business knowledge unit group, a structured reorganization mapping relationship including a step layer, a unit group layer, and a collaborative dimension layer is constructed.

[0020] According to the hierarchical order of the structured reorganization mapping relationship, multiple business knowledge units are arranged and integrated to obtain the structured reorganization result.

[0021] In one embodiment, the step of performing knowledge association analysis on multiple business knowledge units to obtain the logical mapping relationship between the multiple business knowledge units includes:

[0022] For any two business knowledge units under different target power knowledge graph types, business intersection analysis is performed based on the business feature items of the two business knowledge units to identify shared feature items;

[0023] Taking two target business knowledge units with shared feature terms as cross-graph business knowledge unit pairs, and based on a list of typical scenarios for power production environment inspection business, we analyze the business association scenarios of the two target business knowledge units under the shared feature terms.

[0024] Based on the business association scenario and the data source type of the target power knowledge graph type to which the two target business knowledge units belong, the association type of the shared feature item is determined. The association type includes complementary association type, corroborative association type and extended association type.

[0025] In one embodiment, the step of structurally reorganizing multiple business knowledge units based on the logical mapping relationship to obtain a structured reorganization result includes:

[0026] Based on the feature differences of the data source types to which the two target business knowledge units belong, and the information dimension requirements of the shared feature items, the information supplementary dimensions that the shared feature items respectively undertake in the two target business knowledge units are determined.

[0027] An information dimension coverage network is constructed using the shared feature items as nodes and the information supplementary dimensions as edges.

[0028] In the information dimension coverage network, target shared feature items that are associated with a preset number or more business knowledge units are identified, and the information dimension coverage network is divided into feature item association subnets with each target shared feature item as the center.

[0029] Based on the complementary directions of the information supplementary dimensions in the subnet associated with each feature item, a dimension complementarity hierarchy is constructed, and multiple business knowledge units are structurally reorganized according to the order of the dimension complementarity hierarchy to obtain the structured reorganization result.

[0030] In one embodiment, the step of validating the candidate power knowledge question-answer pairs using a large language model and outputting the target power knowledge question-answer pair includes:

[0031] Construct a set of validation criteria that includes question formulation standards and answer formulation standards;

[0032] The candidate power knowledge question-answer pairs are input into a large language model trained with power domain text;

[0033] Using the large language model and based on the verification criteria set, the question and answer statements of the candidate power knowledge question-answer pairs are verified respectively.

[0034] If the verification passes, the candidate power knowledge question-and-answer pair will be determined as the target power knowledge question-and-answer pair;

[0035] If the verification fails, the fit of the structured reorganization result or the question-and-answer template is adjusted according to the reason for failure fed back by the large language model. The candidate power knowledge question-and-answer pair is then verified using the large language model until a target power knowledge question-and-answer pair that passes the verification is generated.

[0036] Secondly, this application also provides a device for constructing question-answer pairs for a multimodal knowledge base in the power industry, comprising:

[0037] The knowledge graph analysis module is used to filter at least two types of target power knowledge graphs based on the target scenarios of power production environment inspection business, and to construct multiple business knowledge units based on the business information of each type of target power knowledge graph.

[0038] The knowledge association analysis module is used to perform knowledge association analysis on multiple business knowledge units to obtain the logical mapping relationship between the multiple business knowledge units.

[0039] The business knowledge restructuring module is used to perform structured restructuring of multiple business knowledge units based on the logical mapping relationship, obtain structured restructuring results, and adapt the structured restructuring results to the corresponding question and answer templates.

[0040] The knowledge question-answer pair generation module is used to transform the structured reorganization result into candidate power knowledge question-answer pairs based on question-answer sentence templates adapted to each business knowledge unit, and to verify the candidate power knowledge question-answer pairs using a large language model, and output the target power knowledge question-answer pair.

[0041] Thirdly, this application also provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to perform the following steps:

[0042] Select at least two target power knowledge graph types, and construct multiple business knowledge units based on the target scenario information of power production environment inspection business for each target power knowledge graph type;

[0043] Perform knowledge association analysis on multiple business knowledge units to obtain the logical mapping relationship between them.

[0044] Based on the logical mapping relationship, multiple business knowledge units are structurally reorganized to obtain the structural reorganization result, and the corresponding question-and-answer sentence template is adapted for the structural reorganization result;

[0045] Based on the question-and-answer sentence templates adapted to each business knowledge unit, the structured reorganization results are transformed into candidate power knowledge question-and-answer pairs, and the candidate power knowledge question-and-answer pairs are verified using a large language model to output the target power knowledge question-and-answer pairs.

[0046] Fourthly, this application also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, performs the following steps:

[0047] Select at least two target power knowledge graph types, and construct multiple business knowledge units based on the target scenario information of power production environment inspection business for each target power knowledge graph type;

[0048] Perform knowledge association analysis on multiple business knowledge units to obtain the logical mapping relationship between them.

[0049] Based on the logical mapping relationship, multiple business knowledge units are structurally reorganized to obtain the structural reorganization result, and the corresponding question-and-answer sentence template is adapted for the structural reorganization result;

[0050] Based on the question-and-answer sentence templates adapted to each business knowledge unit, the structured reorganization results are transformed into candidate power knowledge question-and-answer pairs, and the candidate power knowledge question-and-answer pairs are verified using a large language model to output the target power knowledge question-and-answer pairs.

[0051] Fifthly, this application also provides a computer program product, including a computer program that, when executed by a processor, performs the following steps:

[0052] Select at least two target power knowledge graph types, and construct multiple business knowledge units based on the target scenario information of power production environment inspection business for each target power knowledge graph type;

[0053] Perform knowledge association analysis on multiple business knowledge units to obtain the logical mapping relationship between them.

[0054] Based on the logical mapping relationship, multiple business knowledge units are structurally reorganized to obtain the structural reorganization result, and the corresponding question-and-answer sentence template is adapted for the structural reorganization result;

[0055] Based on the question-and-answer sentence templates adapted to each business knowledge unit, the structured reorganization results are transformed into candidate power knowledge question-and-answer pairs, and the candidate power knowledge question-and-answer pairs are verified using a large language model to output the target power knowledge question-and-answer pairs.

[0056] The aforementioned power multimodal knowledge base question-answering method, apparatus, computer equipment, computer-readable storage medium, and computer program product firstly select multimodal knowledge graph types based on the target scenario and construct fine-grained business knowledge units to provide a multimodal knowledge foundation adapted to the scenario for subsequent analysis; secondly, by performing business semantic association matching and coverage calculation within the same knowledge graph, highly associated unit pairs are selected and their business process dependencies are analyzed to establish a logical mapping within the same graph, ensuring the internal logical coherence of knowledge units; simultaneously, by performing business feature item intersection analysis and associated scenario matching between different knowledge graphs, shared feature items are identified and their... By establishing cross-database logical mappings based on complementary, corroborative, or extended association types, multimodal data barriers are broken down. Then, based on these mapping relationships, a hierarchical, structured reorganization mapping is constructed according to business process step sequences, unit group affiliation and parallel relationships, and collaborative dimensions. Knowledge units are then systematically reorganized to form a logically clear and structurally complete knowledge system. Next, standardized question-and-answer templates are adapted to the reorganization results, achieving the initial transformation of knowledge into standardized question-and-answer pairs. Finally, a large language model trained with domain knowledge is used to automatically verify and correct candidate question-and-answer pairs based on business standards, ensuring the professional accuracy and logical consistency of the output question-and-answer pairs. Attached Figure Description

[0057] To more clearly illustrate the technical solutions in the embodiments of this application or related technologies, the drawings used in the description of the embodiments of this application or related technologies will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0058] Figure 1 This is a diagram illustrating the application environment of the question-and-answer pair construction method for a power multimodal knowledge base in one embodiment.

[0059] Figure 2 This is a flowchart illustrating the method for constructing question-answer pairs in a power multimodal knowledge base, as shown in one embodiment.

[0060] Figure 3This is a flowchart illustrating the method for constructing question-answer pairs in a multimodal knowledge base for electricity, as shown in another embodiment.

[0061] Figure 4 This is a structural block diagram of a power multimodal knowledge base question-answering pair construction device in one embodiment;

[0062] Figure 5 This is an internal structural diagram of a computer device in one embodiment. Detailed Implementation

[0063] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0064] It should be noted that the terms "first," "second," etc., used in this application can be used to describe various elements, but these elements are not limited by these terms. These terms are only used to distinguish the first element from the second element. The terms "comprising" and "having," and any variations thereof, used in this application, are intended to cover non-exclusive inclusion. The term "multiple" used in this application refers to two or more. The term "and / or" used in this application refers to one of the embodiments, or any combination of multiple embodiments.

[0065] The method for constructing question-answer pairs for a multimodal knowledge base in the power industry provided in this application can be applied to, for example... Figure 1 In the application environment shown, terminal 102 communicates with server 104 via a network. A data storage system can store the data that server 104 needs to process. The data storage system can be integrated onto server 104, or it can be located in the cloud or on another network server.

[0066] The terminal 102 can be, but is not limited to, various personal computers, laptops, smartphones, tablets, IoT devices, and portable wearable devices. IoT devices can include smart speakers, smart TVs, smart air conditioners, smart in-vehicle systems, and projection devices. Portable wearable devices can include smartwatches, smart bracelets, and head-mounted displays. Head-mounted displays can be virtual reality (VR) devices, augmented reality (AR) devices, and smart glasses. The server 104 can be a standalone physical server, a server cluster or distributed system composed of multiple physical servers, or a cloud server providing cloud computing services.

[0067] In one exemplary embodiment, such as Figure 2As shown, a method for constructing question-answer pairs for a multimodal knowledge base in the power industry is provided, and this method is applied to... Figure 1 Taking server 104 as an example, the explanation includes the following steps S202 to S208. Wherein:

[0068] Step S202: Select at least two types of target power knowledge graphs, and construct multiple business knowledge units based on the target scenario information of the power production environment inspection business for each type of target power knowledge graph.

[0069] Specifically, the target power knowledge graph type refers to a knowledge graph category that is oriented towards the needs of power production environment inspection business, can cover various types of power knowledge required in this business scenario, and is constructed by the interconnection of specific types of power multimodal data (images, text, videos, time series).

[0070] Power production environment inspection refers to the remote inspection, monitoring, and recording activities carried out on power production sites and related equipment, such as power plants, substations, and transmission lines, in accordance with preset cycles, routes, and standards, to ensure the normal operation of power production equipment and to promptly detect equipment failures and safety hazards.

[0071] A business knowledge unit refers to a basic unit that is based on business information corresponding to a certain target power knowledge graph type, divided according to a certain knowledge granularity, and can independently express a specific business knowledge content under that target power knowledge graph type. Each business knowledge unit has a clear knowledge theme and boundaries.

[0072] Based on the specific target scenarios of power production environment inspection operations, the core business requirements and necessary data modalities within these scenarios are first clarified. For example, in the transformer oil leakage fault identification scenario, it is necessary to simultaneously rely on equipment appearance images and oil level time-series monitoring data. Based on this, at least two target types matching the above requirements are selected from a predefined set of multimodal knowledge graph types, such as image-text knowledge graphs and image-time-series knowledge graphs.

[0073] Subsequently, for each selected target type, its corresponding multimodal business information is collected and organized: for image-text knowledge graphs, the business information includes image feature descriptions of device appearance and related text judgment criteria; for image-time series knowledge graphs, the business information includes image change features of continuous time series and corresponding time series data.

[0074] Finally, based on the pre-defined knowledge granularity rules, the business information under each target type is broken down and summarized into multiple independent knowledge carriers with clear themes and boundaries, namely business knowledge units.

[0075] For the image-time series power knowledge graph type: Business information collected includes time-series changes in transformer appearance oil stain images (e.g., oil stain diameter 2cm on day 1, 4cm on day 3), oil level time-series data (e.g., oil level 2.5m on day 1, 2.3m on day 3), and standard information on the correlation between oil stains and oil levels (e.g., if the rate of oil stain expansion is synchronized with the rate of oil level decline, oil leakage can be confirmed). Business knowledge units are constructed, including "when the diameter of the transformer oil stain increases from 2cm to 4cm within 3 days and the oil level drops from 2.5m to 2.3m, oil leakage can be confirmed," and "when the transformer oil stain shows no change but the oil level drops abnormally, further investigation is needed to check for hidden oil leakage," etc.

[0076] For the image-text power knowledge graph type: the collected business information includes transformer oil level time-series data (such as daily oil level records), text-based oil level anomaly judgment criteria (such as a daily oil level drop exceeding 0.1m is considered abnormal), and text-based oil leakage fault handling procedures (such as after an abnormal drop in oil level, relevant valves should be closed before checking for leaks). The constructed business knowledge units include "A daily drop in transformer oil level exceeding 0.1m is considered an oil level anomaly, and an oil leakage investigation procedure needs to be initiated" and "After an abnormal drop in transformer oil level, the inlet valve should be closed first before checking for leaks at valve interfaces and pipe connections," etc.

[0077] Step S204: Perform knowledge association analysis on multiple business knowledge units to obtain the logical mapping relationship between the multiple business knowledge units.

[0078] Specifically, knowledge association analysis refers to the process of analyzing the relationships between business knowledge units under various target power knowledge graph types to clarify the interactions, mutual support, or mutual constraints between different business knowledge units by analyzing the relationships between them in terms of content logic, business processes, and data dependencies. Logical mapping relationships refer to the correspondences obtained through knowledge association analysis that clearly reflect the logical connections between business knowledge units. In this embodiment, logical mapping relationships include intra-graph logical mapping relationships between business knowledge units under the same target power knowledge graph type, and cross-graph logical mapping relationships between business knowledge units under different target power knowledge graph types.

[0079] Among them, the same-graph logical mapping relationship refers to the mapping relationship formed between different business knowledge units under the same target power knowledge graph type, based on business logic or knowledge content, such as causal relationship, inclusion relationship, parallel relationship, etc. The cross-graph logical mapping relationship refers to the mapping relationship formed between business knowledge units under different target power knowledge graph types, based on the overall logic or data association of power production environment inspection business, such as complementary relationship, verification relationship, support relationship, etc.

[0080] Specifically, for business knowledge units under the same target power knowledge graph type, the power operation and maintenance system can analyze their sequential order in the business process, the inclusion or subordinate relationship of knowledge content, and the causal relationship in business activities such as fault judgment and status assessment to determine the logical mapping relationship with the graph.

[0081] Step S206: Based on the logical mapping relationship, multiple business knowledge units are restructured in a structured manner to obtain the structured restructuring result, and the corresponding question-and-answer sentence template is adapted to the structured restructuring result.

[0082] Specifically, structured reorganization refers to the process of reorganizing and arranging the various business knowledge units based on logical mapping relationships, breaking the original boundaries between knowledge units, and integrating the interrelated business knowledge units into a structured knowledge set with a clear logical structure that can fully express the knowledge system of a certain inspection business area according to the order of business logic and the closeness of the relationship between knowledge content. The result is the structured reorganization result.

[0083] Question and answer templates refer to pre-designed templates with fixed sentence structures that can be filled with content based on specific structured reorganization results, according to the question and answer needs in power production environment inspection business (such as fact inquiry, fault diagnosis, operation process, etc.). The template includes a question section and a corresponding answer section framework. The question section clearly defines the direction of the question, and the answer section reserves filling positions corresponding to the structured knowledge content.

[0084] Furthermore, the power operation and maintenance system integrates various business knowledge units (including business knowledge units under different target power knowledge graph types), removes duplicate knowledge content, and supplements missing knowledge content to obtain a structured reorganization result.

[0085] Step S208: Based on the question-and-answer sentence templates adapted to each business knowledge unit, the structured reorganization results are transformed into candidate power knowledge question-and-answer pairs, and the candidate power knowledge question-and-answer pairs are verified using a large language model to output the target power knowledge question-and-answer pairs.

[0086] Specifically, based on question-and-answer templates with fixed sentence structures adapted for each business knowledge unit, the specific knowledge content corresponding to that unit from the structured reorganization results is filled into the preset question and answer positions in the template, thereby generating candidate power knowledge question-and-answer pairs containing complete question and answer statements. Subsequently, the generated candidate question-and-answer pairs are input into a large language model pre-trained using large-scale power domain text data for automated verification. The large language model analyzes and judges the candidate question-and-answer pairs item by item according to a pre-built set of verification criteria, which includes requirements for the completeness and clarity of the question statement elements, and requirements for the accuracy of the business knowledge source, the logical completeness of the content, and the correctness of the data in the answer statement. If a candidate question-answer pair meets all the verification criteria, it is determined as the final output target power knowledge question-answer pair; if the verification fails, the specific reasons for failure are obtained from the feedback of the large language model (such as ambiguity caused by missing equipment models in the question, incorrect standard sources cited in the answer, etc.). Based on this, the previous structured reorganization results or question-answer sentence template adaptation are adjusted and optimized, and the transformation and verification process is re-executed until the target power knowledge question-answer pair that meets the preset quality standards is produced.

[0087] In the aforementioned method for constructing question-and-answer pairs for a multimodal knowledge base in the power industry, firstly, multimodal knowledge graph types are selected based on the target scenario, and fine-grained business knowledge units are constructed to provide a multimodal knowledge foundation adapted to the scenario for subsequent analysis; secondly, by performing business semantic association matching and coverage calculation within the same knowledge graph, highly associated unit pairs are selected and their business process dependencies are analyzed to establish a logical mapping within the same graph, ensuring the internal logical coherence of knowledge units; simultaneously, by performing business feature item intersection analysis and associated scenario matching between different knowledge graphs, shared feature items are identified, and their complementary, corroborative, or extended association types are determined. First, a cross-database logical mapping is established to break down multimodal data barriers. Then, based on the above mapping relationship, a hierarchical structured reorganization mapping is constructed according to the business process step sequence, unit group affiliation and parallel relationship, and collaboration dimension. Based on this, knowledge units are systematically reorganized to form a logically clear and structurally complete knowledge system. Next, standardized question-and-answer sentence templates are adapted to the reorganization results to achieve the initial transformation of knowledge into standardized question-and-answer pairs. Finally, a large language model trained with domain knowledge is used to automatically verify and correct candidate question-and-answer pairs based on business standards to ensure the professional accuracy and logical consistency of the output question-and-answer pairs.

[0088] In one embodiment, such as Figure 3 As shown, knowledge association analysis is performed on multiple business knowledge units to obtain the logical mapping relationships between them, including:

[0089] Step S302: For any two business knowledge units under the same target power knowledge graph type, based on the pre-built business association semantic dictionary, perform business semantic association matching on the business feature items of the two business knowledge units to obtain the association matching result;

[0090] Step S304: Based on the association matching results, two first business knowledge units that have business semantic association are selected as initial association candidate pairs. The number of business feature items that have business semantic association between the two first business knowledge units is divided by the total number of business feature items between the two first business knowledge units to obtain the feature association coverage rate.

[0091] Step S306: Select initial association candidate pairs with feature association coverage rates not less than a preset coverage rate threshold, and use them as target association candidate pairs;

[0092] Step S308: Based on the business process of power production environment inspection business, analyze the business process dependency relationship between two business knowledge units in the target association candidate pair. The business process dependency relationship includes pre-dependency relationship, post-dependency relationship and parallel relationship.

[0093] Specifically, firstly, business semantic association matching is performed. For any two business knowledge units under this graph type, the business feature items carried by each unit are extracted. These business feature items indicate the specific function of the unit in the inspection business (e.g., "oil leak fault determination function"). Subsequently, a pre-built business association semantic dictionary is invoked. This dictionary defines the semantic association relationships corresponding to standard business processes, equipment specifications, and data interaction rules in the power inspection field. By comparing the business feature items of two units with the association rules in the dictionary, it is determined whether there is a business semantic association between them, and the specific feature item pairs that form the association are recorded, thereby obtaining the association matching result.

[0094] Secondly, the feature association coverage rate is calculated, and highly associated unit pairs are selected. For two business knowledge units that show semantic association according to the association matching results (defined as initial association candidate pairs), their degree of association is quantified. The specific number of feature items with business semantic association between the pair of units is counted, and the total number of business feature items contained in the two units is also calculated. The former is divided by the latter to obtain the feature association coverage rate, which reflects the breadth and depth of association at the business function level. By setting a coverage rate threshold, initial association candidate pairs with a feature association coverage rate not lower than the threshold are selected and identified as target association candidate pairs, thereby focusing on knowledge unit combinations that are highly synergistic in business logic.

[0095] Finally, business process dependencies are analyzed. For each target-related candidate pair, it is analyzed within the standard business process framework of power production environment inspection. Based on the specific business operations corresponding to the two business knowledge units, their execution order and dependencies in the business process are analyzed: if the operation of one unit can only be executed after the operation of another unit is completed, it is defined as a post-dependency relationship. For example, if the fault determination operation is completed after the image analysis operation, the business knowledge unit corresponding to the fault determination operation and the business knowledge unit corresponding to the image analysis operation have a post-dependency relationship. Conversely, it is a pre-dependency relationship. For example, if the image analysis operation needs to be completed before the fault determination operation, the business knowledge unit corresponding to the image analysis operation and the business knowledge unit corresponding to the fault determination operation have a pre-dependency relationship. If the operations of two units can be executed independently or simultaneously without waiting for each other's results, it is defined as a parallel relationship. For example, the image analysis operation of the left surface of the transformer and the image analysis operation of the right surface of the transformer can be performed simultaneously. Through this step, the static semantic associations between units are transformed into dynamic logical mapping relationships that conform to the actual operation process, providing a core basis for subsequent knowledge structure reorganization.

[0096] In this embodiment, by screening multimodal knowledge graphs and constructing business knowledge units, a foundation for multi-source data fusion is laid. Through in-depth correlation analysis of units across and within the same graph, mapping relationships revealing the inherent business logic are formed. Based on these mapping relationships, a hierarchical and structured reorganization is performed to construct a systematic knowledge system. Finally, by adapting question-and-answer templates and validating with a large language model, the structured knowledge is transformed into accurate and consistent standardized question-and-answer pairs. This effectively solves the problems of low defect identification rate in single-modal analysis, one-sided status evaluation, and poor efficiency and consistency in maintenance plan generation, thereby improving the accuracy and intelligence level of power operation and maintenance knowledge application.

[0097] In one embodiment, multiple business knowledge units are structurally reorganized based on logical mapping relationships to obtain a structured reorganization result, including:

[0098] Determine the corresponding business process steps in the business process for all pre-dependencies and post-dependencies in the logical mapping relationship, and sort each business process step in the execution order to obtain the business process step sequence;

[0099] Based on the attribution relationship between multiple business knowledge units and each business process step in the business process step sequence, the attribution business knowledge unit group corresponding to each business process step is determined, and based on the parallel relationship, the parallel business knowledge unit group that is executed in parallel within each attribution business knowledge unit group is determined.

[0100] Based on the business process step sequence, the belonging business knowledge unit group, and the collaborative operation dimension of business knowledge units within the parallel business knowledge unit group, a structured reorganization mapping relationship including the step layer, unit group layer, and collaborative dimension layer is constructed.

[0101] According to the hierarchical order of the structured reorganization mapping relationship, multiple business knowledge units are arranged and integrated to obtain the structured reorganization result.

[0102] Specifically, firstly, based on the pre- and post-dependencies in the logical mapping relationship, the specific steps corresponding to these dependencies in the standard business process are identified, and these steps are defined as business process steps. These steps are then ordered according to the order of business execution, forming a sequence of business process steps from start to finish, thereby establishing the time-based logical thread of the reorganization process.

[0103] Secondly, based on the core business operations of each business knowledge unit, determine which business process step in the above sequence it belongs to. Define the set of all business knowledge units belonging to the same business process step as a group of affiliated business knowledge units. Then, within a group, based on the parallel relationships in the logical mapping, further divide business knowledge units that can execute operations simultaneously into smaller parallel business knowledge unit groups, thereby clarifying the parallel execution structure within the step.

[0104] Then, a multi-level structured reorganization mapping relationship is constructed. This mapping uses a step layer consisting of a sequence of business process steps as the top-level framework; below, each step is associated with a corresponding business knowledge unit group, forming a unit group layer; finally, under each group, a collaborative operation dimension layer is refined based on the specific operational perspectives of each unit within its parallel business knowledge unit groups (such as for different parts of the device or different data dimensions). These three levels together constitute a knowledge organization blueprint from macro-process to micro-operation.

[0105] Finally, based on the constructed structured reorganization mapping relationship, and following the hierarchical order of "step layer → unit group layer → collaborative operation dimension layer," all business knowledge units are systematically arranged, integrated, and their content is connected to generate the final structured reorganization result. This result transforms the originally scattered multimodal knowledge into a hierarchical and logically coherent organic whole, laying a solid data foundation for generating high-quality question-and-answer pairs that conform to business logic.

[0106] For example, the structured reconstruction result is as follows: Step 1: Transformer image acquisition. Unit group 1: Simultaneously perform the following image acquisition and feature extraction operations: (1) Valve interface image acquisition dimension: acquire the valve interface image of transformer S11-2500kVA and extract oil trace features; (2) Surface image acquisition dimension: acquire the surface image of transformer and extract oil trace features. Step 3: Oil leakage fault judgment. Unit group 2: Valve interface oil leakage judgment dimension: the oil trace diameter at the valve interface of transformer S11-2500kVA exceeds 5cm, which is an oil leakage fault. Step 4: Normal appearance judgment. Unit group 3: Normal appearance judgment dimension: the transformer surface without dark and shiny oil traces is a normal appearance state. Step 5: Result integration. Unit group 4: Judgment result integration dimension: integrate the oil leakage fault judgment result and the normal appearance judgment result to form the final inspection conclusion.

[0107] In this embodiment, by transforming logical mapping relationships into a sequence of business process steps, and establishing affiliated and parallel business knowledge unit groups accordingly, a multi-layered reorganization mapping of step-unit group-collaboration dimensions is constructed. This integrates discrete, heterogeneous, multimodal business knowledge units into a structured knowledge system that is hierarchical, logically rigorous, and faithfully reflects the actual operational process. This solves the problems of fragmented multi-source knowledge and weak logical connections, providing a direct and reliable structured data foundation for generating high-quality, standardized question-and-answer pairs that conform to business specifications and are context-coherent.

[0108] In one embodiment, knowledge association analysis is performed on multiple business knowledge units to obtain logical mapping relationships between the multiple business knowledge units, including:

[0109] For any two business knowledge units under different target power knowledge graph types, business intersection analysis is performed based on the business feature items of the two business knowledge units to identify shared feature items;

[0110] Taking two target business knowledge units with shared feature items as cross-graph business knowledge unit pairs, and based on the typical scenario list of power production environment inspection business, we analyze the business association scenarios of the two target business knowledge units under shared feature items;

[0111] Based on the business association scenario and the data source type of the target power knowledge graph to which the two target business knowledge units belong, the association type of shared feature items is determined. The association types include complementary association types, corroborative association types, and extended association types.

[0112] Specifically, firstly, a cross-graph business intersection analysis is performed. For any two business knowledge units belonging to different target power knowledge graph types, the core attributes representing their business functions, i.e., business feature items, are extracted. By comparing the business feature item sets of the two units, feature items that exist in both sets are identified and defined as shared feature items. These shared feature items reveal the common concerns of different modal knowledge units in terms of business functions and serve as the logical link for constructing cross-graph connections.

[0113] Secondly, business-related scenarios are identified. A combination of two business knowledge units sharing common features is defined as a cross-graph business knowledge unit pair. Then, based on a pre-compiled list of typical scenarios for power production environment inspections, which lists standardized business scenarios such as "accurate determination of transformer oil leakage faults" and "investigation of abnormal oil levels" and their corresponding multimodal operation combinations, the core operations of the two units in this cross-graph business knowledge unit pair are analyzed to determine which specific scenario in the list they can jointly support or apply to. This clarifies the specific business context in which the pair of units exerts its synergistic effect, i.e., the business-related scenario.

[0114] Finally, the association type of the shared feature item is determined. Combining the business association scenario determined in the previous step, and the data source types corresponding to the target power knowledge graph types to which the two units in this cross-graph business knowledge unit belong (e.g., "image + text", "image + time series", "time series + text"), a comprehensive analysis is conducted on the role patterns and logical relationships of the two data sources in describing or supporting the same shared feature item: if the two data sources provide information from different dimensions, and their combination is necessary to fully describe the feature item, it is determined to be a "complementary association type"; if the two data sources can mutually verify the same conclusion, it is determined to be a "confirmatory association type"; if the analysis of one data source forms the basis for in-depth analysis of another data source, it is determined to be an "extended association type". This step ultimately transforms the static cross-graph association into a cross-graph logical mapping relationship with clear semantics and business logic.

[0115] In this embodiment, shared feature items are located through cross-database business intersection analysis, constructing the link between heterogeneous knowledge units. A list of typical scenarios clarifies the specific business context of their collaborative applications, ensuring that the relationships align with actual operation and maintenance needs. Finally, based on the data source type, the association types (complementary, corroborative, or extended) are determined, transforming the static coexistence of cross-modal data into dynamic collaborative relationships with clear semantics and logical guidance. This effectively solves the problem of fragmented and difficult-to-integrate multi-source heterogeneous data, laying the foundation for building a unified and comprehensive knowledge system for power equipment condition assessment and decision support.

[0116] In one embodiment, multiple business knowledge units are structurally reorganized based on logical mapping relationships to obtain a structured reorganization result, including:

[0117] Based on the characteristic differences of the data source types of the two target business knowledge units and the information dimension requirements of the shared feature items, the information supplementation dimensions that the shared feature items undertake in the two target business knowledge units are determined respectively.

[0118] Using shared feature items as nodes and supplementary information dimensions as edges, an information dimension coverage network is constructed.

[0119] In the information dimension coverage network, target shared feature items that are associated with a preset number or more of business knowledge units are identified, and the information dimension coverage network is divided into feature item association subnets with each target shared feature item as the center.

[0120] Based on the complementary directions of the information supplementary dimensions in the subnet associated with each feature item, a dimension complementarity hierarchy is constructed, and multiple business knowledge units are structurally reorganized according to the order of the dimension complementarity hierarchy to obtain the structured reorganization result.

[0121] Specifically, firstly, the specific dimensions of information complementarity are determined. For cross-graph knowledge unit pairs with complementary relationships, the essential differences in the data source types of the two units are analyzed. For example, the static spatial description and standard definition capabilities of the "image + text" data source are compared with the dynamic trend characterization capabilities of the "image + time series" data source. Considering the complete information dimension requirements that the shared feature item (such as "oil leak fault determination") needs to cover in business (e.g., simultaneous static threshold and dynamic trend), the information supplementation role of each unit in the description of the shared feature item is clarified. For example, one unit provides static space and standard dimensions, while the other unit provides dynamic time and trend dimensions. For instance, two second-target knowledge units form a bidirectional information complementarity relationship on the shared feature item. Specifically, unit A (image-text power knowledge graph) supplements the shared feature item F1 of unit G (image-time series power knowledge graph) with standard basis and static feature information through static spatial dimensions and standard basis dimensions; unit G supplements the shared feature item F1 of unit A with trend change and dynamic feature information through dynamic time dimensions and trend analysis dimensions. Together, they improve the oil leak fault determination knowledge of F1.

[0122] Secondly, a feature-centric association network is constructed and divided. Using shared feature items as core nodes and their supplementary information dimensions within various units as directed edges pointing to those nodes, an information dimension coverage network is built. Within this global network, pivotal feature items that are associated with multiple (at a preset number, such as two or more) different business knowledge units are identified and designated as target shared feature items. Subsequently, with each target shared feature item as the center, all directly connected business knowledge unit nodes and their connecting edges are extracted, forming an independent feature item association subnet.

[0123] Finally, a hierarchical reorganization is performed based on complementary directions. Within each feature-related subnet, according to the direction of provision from knowledge units to shared feature items indicated by the information supplement dimension edges, and combined with the sequence of business logic (e.g., standards usually precede data collection, and data collection precedes trend analysis), the associated knowledge units are divided into complementary dimensional levels with sequential or supporting relationships, such as "basic standards layer → data collection layer → analysis trigger layer". Ultimately, according to the constructed complementary dimensional hierarchical order, all involved business knowledge units are logically arranged and their content integrated to generate a systematic and structured reorganization result, thereby ensuring that multimodal knowledge around the core feature items is integrated in an orderly and complete manner.

[0124] In this embodiment, complementary and related cross-graph knowledge units are mapped to explicit information supplementary dimensions, and an information dimension coverage network is constructed with shared feature items as the core to identify pivotal target shared feature items and their associated subnets. Furthermore, a dimension complementarity hierarchy is established based on business logic and complementary directions, ultimately achieving the orderly integration of multi-source heterogeneous knowledge around core business features. This effectively solves the problems of loose information and unclear hierarchy in cross-modal data fusion, ensuring that the reorganized knowledge system possesses both structure and integrity.

[0125] In one embodiment, a large language model is used to validate candidate power knowledge question-answer pairs, and a target power knowledge question-answer pair is output, including:

[0126] Construct a set of validation criteria that includes question formulation standards and answer formulation standards;

[0127] The candidate power knowledge question-answer pairs are input into a large language model trained with power domain text;

[0128] Using a large language model and based on a set of validation criteria, the question and answer statements of candidate power knowledge question-answer pairs are validated respectively.

[0129] If the verification is successful, the candidate power knowledge question-and-answer pair will be determined as the target power knowledge question-and-answer pair;

[0130] If the verification fails, the fit of the structured recombination result or question-and-answer template is adjusted based on the reasons for failure provided by the large language model. The candidate power knowledge question-and-answer pairs are then verified using the large language model until the target power knowledge question-and-answer pair that passes the verification is generated.

[0131] Specifically, firstly, a set of verification standards with clear objectives is constructed. This set of standards comprises two parts: problem formulation standards and answer formulation standards. Problem formulation standards constrain the standardization of the questions posed, requiring them to clearly include key business elements such as equipment model, fault type, and business scenario, without ambiguity. Answer formulation standards, on the other hand, constrain the accuracy and completeness of the answers, requiring them to accurately cite original business knowledge sources, fully cover relevant business knowledge modules, strictly follow the correct business logic order, and ensure all technical terms and data thresholds are accurate.

[0132] Secondly, automated verification based on a domain-specific knowledge model is performed. Candidate power knowledge question-answer pairs are input into a large language model pre-trained on massive amounts of power-related text data (such as equipment manuals, fault cases, operating procedures, and industry standards). This model intelligently analyzes the candidate question-answer pairs based on the aforementioned verification criteria set: verifying whether the question clearly points to specific business needs; checking whether the knowledge sources in the answers are consistent with the original information used to construct the business knowledge units; and reviewing whether the knowledge content is complete, logically consistent, and the data accurate.

[0133] Finally, the closed-loop optimization of the knowledge system is driven by the verification results. If a candidate question-answer pair passes all verifications, it is directly output as the target power knowledge question-answer pair. If the verification fails, the specific reasons for failure are received from the large language model (e.g., "key knowledge sources are missing in the answer" or "the question lacks equipment models, leading to ambiguity"). Based on this feedback, the previous structured reorganization results or the adaptation process of the question-answer sentence template are reviewed and adjusted. After correcting the defects, candidate question-answer pairs are regenerated and submitted for verification again. This process is iterated until high-quality target power knowledge question-answer pairs that fully meet the verification criteria are generated, thereby ensuring the reliability and professionalism of the operation and maintenance knowledge output.

[0134] In this embodiment, by constructing a specialized set of verification criteria and utilizing a large language model trained with domain knowledge for automated verification, multi-dimensional intelligent review of candidate question-answer pairs is achieved in terms of the completeness of business elements, the accuracy of knowledge sources, logical rigor, and data precision. When verification fails, the preceding knowledge reorganization or template adaptation stages are precisely backtracked and adjusted based on the specific reasons fed back by the model, forming a closed-loop iterative mechanism of "generation-verification-optimization." This effectively solves the problems of low efficiency and inconsistent quality of manual review, as well as the difficulty of traditional methods in ensuring the professional accuracy of knowledge output.

[0135] In one embodiment, knowledge association analysis is performed on multiple business knowledge units to obtain logical mapping relationships between the multiple business knowledge units, including:

[0136] For each cross-graph business knowledge unit pair, if the association type of the shared feature item is a complementary association type, then the data source type is compared based on the data source type of the target power knowledge graph to which the two second target business knowledge units belong, and the differences in data source features are obtained.

[0137] Based on the differences in data source characteristics between the two target business knowledge units and the feature item information of the power production environment inspection business, the information supplementation dimension undertaken by the shared feature item in the two target business knowledge units is determined; the feature item information indicates the information dimension covered by each business feature item.

[0138] Using shared feature items as nodes and the information supplementation dimensions undertaken by the shared feature items in two target business knowledge units as edges, an information dimension coverage network is constructed to obtain the cross-graph logical mapping relationship between various business knowledge units.

[0139] In one embodiment, knowledge association analysis is performed on multiple business knowledge units to obtain logical mapping relationships between the multiple business knowledge units, including:

[0140] For each cross-graph business knowledge unit pair, if the association type of the shared feature item is the corroboration association type, then the shared feature item is analyzed based on the business feature items of the two second target business knowledge units to obtain the feature item description perspectives of the two second target business knowledge units for the shared feature item; the feature item description perspective indicates the dimension by which the business knowledge unit describes the attributes or status of the shared feature item in the power production environment inspection business based on the data source type of its own target power knowledge graph type.

[0141] Based on two second-target business knowledge units, the similarity and difference analysis of the feature item description perspectives of shared feature items is carried out to obtain common description perspectives and difference description perspectives.

[0142] Using shared feature items as main nodes, common description perspectives as connecting edges, and different description perspectives as supplementary nodes, a multi-perspective verification network is constructed to obtain cross-graph logical mapping relationships between various business knowledge units.

[0143] In one embodiment, knowledge association analysis is performed on multiple business knowledge units to obtain logical mapping relationships between the multiple business knowledge units, including:

[0144] For each cross-graph business knowledge unit pair, if the shared feature item association type is an extended association type, then the business knowledge extension direction and extension path are determined based on the data source complexity of the data source type of the target power knowledge graph to which the two second target business knowledge units belong.

[0145] Based on the shared feature terms, the business feature terms of the two second target business knowledge units are traversed along the extension path to determine the common extension starting point of the two second target business knowledge units, as well as the extension intermediate point and extension endpoint of the two second target business knowledge units respectively.

[0146] Using a common starting point and the intermediate and endpoint points of extension on the two second target business knowledge units as information extension levels, information extension chains are constructed in the order of extension directions to obtain cross-graph logical mapping relationships between various business knowledge units.

[0147] It should be understood that although the steps in the flowcharts of the embodiments described above are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowcharts of the embodiments described above may include multiple steps or multiple stages. These steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the steps or stages of other steps.

[0148] Based on the same inventive concept, this application also provides a power multimodal knowledge base question-answering pair construction device for implementing the above-mentioned power multimodal knowledge base question-answering pair construction method. The solution provided by this device is similar to the solution described in the above method. Therefore, the specific limitations of one or more power multimodal knowledge base question-answering pair construction device embodiments provided below can be found in the limitations of the power multimodal knowledge base question-answering pair construction method above, and will not be repeated here.

[0149] In one exemplary embodiment, such as Figure 4 As shown, a device for constructing question-answer pairs for a multimodal knowledge base in the power industry is provided, comprising:

[0150] The knowledge graph analysis module 402 is used to filter at least two types of target power knowledge graphs based on the target scenario of power production environment inspection business, and to construct multiple business knowledge units based on the business information of each type of target power knowledge graph.

[0151] The knowledge association analysis module 404 is used to perform knowledge association analysis on multiple business knowledge units to obtain the logical mapping relationship between the multiple business knowledge units.

[0152] The business knowledge restructuring module 406 is used to restructure multiple business knowledge units in a structured manner based on logical mapping relationships, obtain structured restructuring results, and adapt the structured restructuring results to corresponding question and answer templates.

[0153] The knowledge question and answer pair generation module 408 is used to transform the structured reorganization results into candidate power knowledge question and answer pairs based on the question and answer sentence templates adapted to each business knowledge unit, and to verify the candidate power knowledge question and answer pairs using a large language model, and output the target power knowledge question and answer pair.

[0154] In an exemplary embodiment, the knowledge association analysis module 404 is specifically used to perform business semantic association matching on the business feature items of any two business knowledge units under the same target power knowledge graph type, based on a pre-built business association semantic dictionary, to obtain association matching results; based on the association matching results, the two first business knowledge units with business semantic association are selected as initial association candidate pairs; the number of business feature items with business semantic association between the two first business knowledge units is divided by the total number of business feature items of the two first business knowledge units to obtain the feature association coverage rate; the initial association candidate pairs with feature association coverage rates not less than a preset coverage rate threshold are selected as target association candidate pairs; based on the business process of power production environment inspection business, the business process dependency relationship between the two business knowledge units in the target association candidate pair is analyzed, and the business process dependency relationship includes pre-dependency relationship, post-dependency relationship and parallel relationship.

[0155] In an exemplary embodiment, the business knowledge reorganization module 406 is specifically used to determine the business process steps corresponding to all pre-dependencies and post-dependencies in the logical mapping relationship in the business process, and sort each business process step according to the execution order to obtain a business process step sequence; based on the attribution relationship between multiple business knowledge units and each business process step in the business process step sequence, determine the attribution business knowledge unit group corresponding to each business process step, and based on the parallel relationship, determine the parallel business knowledge unit groups that are executed in parallel within each attribution business knowledge unit group; based on the business process step sequence, the attribution business knowledge unit group, and the collaborative operation dimension of the business knowledge units within the parallel business knowledge unit group, construct a structured reorganization mapping relationship including a step layer, a unit group layer, and a collaborative dimension layer; and arrange and integrate multiple business knowledge units according to the hierarchical order of the structured reorganization mapping relationship to obtain a structured reorganization result.

[0156] In an exemplary embodiment, the knowledge association analysis module 404 is specifically used to perform business intersection analysis on any two business knowledge units under different target power knowledge graph types, based on the business feature items of the two business knowledge units, to identify shared feature items; taking the two target business knowledge units with shared feature items as cross-graph business knowledge unit pairs, and based on the typical scenario list of power production environment inspection business, to analyze the business association scenarios of the two target business knowledge units under shared feature items; and based on the business association scenarios and the data source type of the target power knowledge graph type to which the two target business knowledge units belong, to determine the association type of the shared feature items, including complementary association type, corroborative association type, and extended association type.

[0157] In an exemplary embodiment, the business knowledge reorganization module 406 is specifically used to determine the information supplementary dimensions that the shared feature items undertake in the two target business knowledge units based on the feature differences of the data source types to which the two target business knowledge units belong and the information dimension requirements of the shared feature items; to construct an information dimension coverage network with the shared feature items as nodes and the information supplementary dimensions as edges; to determine the target shared feature items that are associated with more than a preset number of business knowledge units in the information dimension coverage network, and to divide the information dimension coverage network into feature item association subnets with each target shared feature item as the center; to construct a dimension complementarity hierarchy based on the complementary direction of the information supplementary dimensions in each feature item association subnet, and to perform structured reorganization of multiple business knowledge units in accordance with the order of the dimension complementarity hierarchy to obtain the structured reorganization result.

[0158] In an exemplary embodiment, the knowledge question-answer pair generation module 408 is used to construct a verification standard set containing question expression standards and answer expression standards; input candidate power knowledge question-answer pairs into a large language model trained on power domain text; verify the question expression and answer expression of the candidate power knowledge question-answer pairs respectively through the large language model and according to the verification standard set; if the verification passes, the candidate power knowledge question-answer pair is determined as the target power knowledge question-answer pair; if the verification fails, the fit of the structured recombination result or question-answer sentence template is adjusted according to the reason for failure fed back by the large language model, and the candidate power knowledge question-answer pairs are verified using the large language model until the target power knowledge question-answer pair that passes the verification is generated.

[0159] The various modules in the aforementioned power multimodal knowledge base question-and-answer pair construction device can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in the processor of a computer device in hardware form or independent of it, or stored in the memory of a computer device in software form, so that the processor can call and execute the operations corresponding to each module.

[0160] In one exemplary embodiment, a computer device is provided, which may be a server, and its internal structure diagram may be as follows: Figure 5 As shown, the computer device includes a processor, memory, input / output (I / O) interfaces, and a communication interface. The processor, memory, and I / O interfaces are connected via a system bus, and the communication interface is also connected to the system bus via the I / O interfaces. The processor provides computational and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system, computer programs, and a database. The internal memory provides the environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The database stores structured reconstruction results data. The I / O interfaces are used for exchanging information between the processor and external devices. The communication interface is used for communication with external terminals via a network connection. When executed by the processor, the computer program implements a method for constructing question-answer pairs for a multimodal knowledge base in the power industry.

[0161] Those skilled in the art will understand that Figure 5 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0162] In one exemplary embodiment, a computer device is provided, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps of the method described above.

[0163] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon, which, when executed by a processor, implements the steps of the above-described method.

[0164] In one embodiment, a computer program product is provided, including a computer program that, when executed by a processor, implements the steps of the method described above.

[0165] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of the relevant data must comply with relevant regulations.

[0166] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. Any references to memory, databases, or other media used in the embodiments provided in this application can include at least one of non-volatile memory and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM can take many forms, such as Static Random Access Memory (SRAM) or Dynamic Random Access Memory (DRAM). The databases involved in the embodiments provided in this application may include at least one type of relational database and non-relational database. Non-relational databases may include, but are not limited to, blockchain-based distributed databases. The processors involved in the embodiments provided in this application may be general-purpose processors, central processing units, graphics processing units, digital signal processors, programmable logic devices, quantum computing-based data processing logic devices, artificial intelligence (AI) processors, etc., and are not limited to these.

[0167] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this application.

[0168] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of this patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this application should be determined by the appended claims.

Claims

1. A method for constructing question-answer pairs in a multimodal knowledge base for electricity, characterized in that, The method includes: Select at least two target power knowledge graph types, and construct multiple business knowledge units based on the target scenario information of power production environment inspection business for each target power knowledge graph type; Perform knowledge association analysis on multiple business knowledge units to obtain the logical mapping relationship between them. Based on the logical mapping relationship, multiple business knowledge units are structurally reorganized to obtain the structural reorganization result, and the corresponding question-and-answer sentence template is adapted for the structural reorganization result; Based on the question-and-answer sentence templates adapted to each business knowledge unit, the structured reorganization results are transformed into candidate power knowledge question-and-answer pairs, and the candidate power knowledge question-and-answer pairs are verified using a large language model to output the target power knowledge question-and-answer pairs.

2. The method according to claim 1, characterized in that, The step of performing knowledge association analysis on multiple business knowledge units to obtain the logical mapping relationship between the multiple business knowledge units includes: For any two business knowledge units under the same target power knowledge graph type, based on the pre-built business association semantic dictionary, business semantic association matching is performed on the business feature items of the two business knowledge units to obtain the association matching result. Based on the association matching results, two first business knowledge units that have business semantic association are selected as initial association candidate pairs. The number of business feature items that have business semantic association between the two first business knowledge units is divided by the total number of business feature items between the two first business knowledge units to obtain the feature association coverage rate. Select initial candidate pairs of associations with a feature association coverage rate not less than a preset coverage rate threshold, and use them as target candidate pairs of associations; Based on the business process of power production environment inspection, the business process dependency relationship between two business knowledge units in the target association candidate pair is analyzed. The business process dependency relationship includes pre-dependency relationship, post-dependency relationship and parallel relationship.

3. The method according to claim 2, characterized in that, The step of structurally reorganizing multiple business knowledge units based on the logical mapping relationship to obtain a structured reorganization result includes: Determine the business process steps corresponding to all pre-dependencies and post-dependencies in the logical mapping relationship in the business process, and sort each business process step in the execution order to obtain a business process step sequence. Based on the attribution relationship between multiple business knowledge units and each business process step in the business process step sequence, the attribution business knowledge unit group corresponding to each business process step is determined, and based on the parallel relationship, the parallel business knowledge unit group that is executed in parallel within each attribution business knowledge unit group is determined. Based on the business process step sequence, the belonging business knowledge unit group, and the collaborative operation dimension of business knowledge units within the parallel business knowledge unit group, a structured reorganization mapping relationship including a step layer, a unit group layer, and a collaborative dimension layer is constructed. According to the hierarchical order of the structured reorganization mapping relationship, multiple business knowledge units are arranged and integrated to obtain the structured reorganization result.

4. The method according to claim 1, characterized in that, The step of performing knowledge association analysis on multiple business knowledge units to obtain the logical mapping relationship between the multiple business knowledge units includes: For any two business knowledge units under different target power knowledge graph types, business intersection analysis is performed based on the business feature items of the two business knowledge units to identify shared feature items; Taking two target business knowledge units with shared feature terms as cross-graph business knowledge unit pairs, and based on a list of typical scenarios for power production environment inspection business, we analyze the business association scenarios of the two target business knowledge units under the shared feature terms. Based on the business association scenario and the data source type of the target power knowledge graph type to which the two target business knowledge units belong, the association type of the shared feature item is determined. The association type includes complementary association type, corroborative association type and extended association type.

5. The method according to claim 4, characterized in that, The step of structurally reorganizing multiple business knowledge units based on the logical mapping relationship to obtain a structured reorganization result includes: Based on the feature differences of the data source types to which the two target business knowledge units belong, and the information dimension requirements of the shared feature items, the information supplementary dimensions that the shared feature items respectively undertake in the two target business knowledge units are determined. An information dimension coverage network is constructed using the shared feature items as nodes and the information supplementary dimensions as edges. In the information dimension coverage network, target shared feature items that are associated with a preset number or more business knowledge units are identified, and the information dimension coverage network is divided into feature item association subnets with each target shared feature item as the center. Based on the complementary directions of the information supplementary dimensions in the subnet associated with each feature item, a dimension complementarity hierarchy is constructed, and multiple business knowledge units are structurally reorganized according to the order of the dimension complementarity hierarchy to obtain the structured reorganization result.

6. The method according to claim 1, characterized in that, The process of using a large language model to validate the candidate power knowledge question-answer pairs and outputting the target power knowledge question-answer pairs includes: Construct a set of validation criteria that includes question formulation standards and answer formulation standards; The candidate power knowledge question-answer pairs are input into a large language model trained with power domain text; Using the large language model and based on the verification criteria set, the question and answer statements of the candidate power knowledge question-answer pairs are verified respectively. If the verification passes, the candidate power knowledge question-and-answer pair will be determined as the target power knowledge question-and-answer pair; If the verification fails, the fit of the structured reorganization result or the question-and-answer template is adjusted according to the reason for failure fed back by the large language model. The candidate power knowledge question-and-answer pair is then verified using the large language model until a target power knowledge question-and-answer pair that passes the verification is generated.

7. A device for constructing question-answer pairs for a multimodal knowledge base in the power industry, characterized in that, The device includes: The knowledge graph analysis module is used to filter at least two types of target power knowledge graphs based on the target scenarios of power production environment inspection business, and to construct multiple business knowledge units based on the business information of each type of target power knowledge graph. The knowledge association analysis module is used to perform knowledge association analysis on multiple business knowledge units to obtain the logical mapping relationship between the multiple business knowledge units. The business knowledge restructuring module is used to perform structured restructuring of multiple business knowledge units based on the logical mapping relationship, obtain structured restructuring results, and adapt the structured restructuring results to the corresponding question and answer templates. The knowledge question-answer pair generation module is used to transform the structured reorganization result into candidate power knowledge question-answer pairs based on question-answer sentence templates adapted to each business knowledge unit, and to verify the candidate power knowledge question-answer pairs using a large language model, and output the target power knowledge question-answer pair.

8. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 6.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 6.

10. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 6.

Images