A knowledge graph-based power repair model fine-tuning method and fine-tuning system

Abstract

This application relates to the field of data processing technology, and in particular to a method and system for fine-tuning a power emergency repair model based on a knowledge graph. The method includes classifying and labeling sample multimodal data, performing categorized data management to obtain effective multimodal data, and constructing a power equipment emergency repair knowledge graph based on this effective data. Candidate models and a fit evaluation system are determined based on demand indicators. The candidate models are then subjected to standardized testing based on the fit evaluation system, and the candidate model with the highest overall fit score is selected as the target base model. Triple data in the power equipment emergency repair knowledge graph is identified, and the triple data is fused with the effective multimodal data to form training sample pairs. The target base model is then fine-tuned hierarchically based on the training sample pairs and a hierarchical training dataset to obtain the fine-tuned model. This application facilitates improved efficiency in operation and maintenance emergency repair work, thereby enhancing the reliability of power supply.

Description

Technical Field

[0001] This application relates to the field of data processing technology, and in particular to a method and system for fine-tuning a power emergency repair model based on knowledge graphs. Background Technology

[0002] During long-term operation, power equipment is susceptible to the combined effects of complex natural environment and high load conditions, which accelerates the aging process of the equipment, resulting in a high failure rate and frequent emergency repairs. This not only significantly increases the difficulty of repair and operation and maintenance costs, but also makes it difficult for some outdated equipment to adapt to the current high load operation requirements of the power grid, seriously restricting the reliability of power supply.

[0003] Currently, power grid operation and maintenance (O&M) still relies heavily on manual inspection, lacking efficient intelligent monitoring and analysis methods. To address this challenge, general-purpose large-scale models have been introduced to improve the intelligence level and overall efficiency of O&M. However, the training data for these models focuses on general knowledge, exhibiting significant knowledge comprehension biases when adapted to the specialized scenario of power O&M. Specifically, traditional model fine-tuning methods fail to fully integrate structured knowledge from the power sector, resulting in insufficient ability to identify and reason about power equipment fault characteristics and repair procedures. This leads to a series of problems, including low accuracy in power equipment fault identification, a lack of targeted repair plan recommendations, and poor adaptation to professional standards, ultimately hindering a genuine improvement in O&M efficiency. Summary of the Invention

[0004] To improve the efficiency of operation, maintenance and emergency repair work, thereby enhancing the reliability of power supply, this application provides a knowledge graph-based method and system for fine-tuning power emergency repair models.

[0005] Firstly, this application provides a method for fine-tuning a power emergency repair model based on a knowledge graph, employing the following technical solution: A method for fine-tuning a power emergency repair model based on knowledge graphs, comprising: Collect sample multimodal data in the field of power grid operation, maintenance and emergency repair, and classify and label the sample multimodal data to obtain classified and labeled sample multimodal data to be processed. The sample multimodal data includes text data, image data and structured data. The sample multimodal unprocessed data is subjected to categorized data management and control processing to obtain effective multimodal data for management and control, and a knowledge graph for emergency repair of power equipment is constructed based on the effective multimodal data for management and control. Identify the demand indicators corresponding to the effective data of the multimodal management and control, and determine the candidate models and the suitability evaluation system based on the demand indicators; Based on the fitness evaluation system, a standardized test is performed on each candidate model to obtain the comprehensive fitness score corresponding to each candidate model, and the candidate model with the highest comprehensive fitness score is used as the target base model. A hierarchical training dataset is constructed based on the power equipment emergency repair knowledge graph. The hierarchical training dataset includes a general power knowledge dataset, a fault diagnosis special dataset, and an emergency repair practice dataset. Identify the triplet data in the knowledge graph of power equipment emergency repair, and perform fusion processing based on the triplet data and the effective data of the multimodal management and control to form training sample pairs; Based on the training sample pairs and the hierarchical training dataset, the target base model is fine-tuned hierarchically to obtain the fine-tuned model.

[0006] By adopting the above technical solutions and constructing a knowledge graph for power equipment emergency repair, structured knowledge such as equipment parameters, fault types, and repair procedures are integrated into fine-tuning, which facilitates the improvement of the fault diagnosis accuracy of the power emergency repair model. Simultaneously, layered training is conducted using general power knowledge, fault diagnosis-specific data, and practical repair exercises to guide the power emergency repair model to gradually master professional capabilities, avoiding knowledge confusion. This enhances the model's ability to distinguish highly similar faults and reason about complex repair scenarios. By fusing triplet data from the power equipment emergency repair knowledge graph with effective multimodal management data to form training sample pairs, and then performing layered fine-tuning of the target basic model based on these training sample pairs and the layered training dataset, the power emergency repair model can simultaneously absorb structured knowledge and multi-source data features. This improves the model's accuracy in processing multimodal data, enabling precise fault diagnosis, thereby increasing the efficiency of operation, maintenance, and emergency repair work, and ultimately enhancing the reliability of power supply.

[0007] In one possible implementation, the construction of a power equipment emergency repair knowledge graph based on the effective multimodal data of the management and control system includes: The core entities in the field of power equipment emergency repair are extracted from the effective data of the multimodal management and control. The core entities include equipment entities, fault entities, emergency repair resource entities, and process entities. The relationships between core entities are determined based on the business logic of the power equipment emergency repair field. A knowledge graph storage architecture is built based on a pre-defined graph database, and a triplet structure with entity-attribute-relationship as the core is constructed to form a knowledge system for emergency repair of power equipment. Establish the association and mapping relationship between the power equipment emergency repair knowledge system and the effective multimodal data of management and control to obtain the power equipment emergency repair knowledge graph.

[0008] By adopting the above technical solution, core entities such as equipment, faults, emergency repair resources, and processes are accurately extracted from effective multimodal data. The relationships between these entities are clarified by combining them with the business logic of power emergency repair. A triplet storage architecture is built based on a pre-set graph database, forming a systematic knowledge system for power equipment emergency repair and establishing a mapping relationship with multimodal data. This constructs a knowledge graph for power equipment emergency repair, facilitating the transformation of scattered multi-source data into structured and systematic professional knowledge. This provides accurate domain knowledge support for the model. Furthermore, the knowledge graph for power equipment emergency repair achieves deep binding between multimodal data and professional knowledge, enabling the model to quickly associate key information such as equipment parameters, fault characteristics, and emergency repair processes, avoiding knowledge fragmentation. This, in turn, improves the model's understanding of professional power emergency repair scenarios, providing a high-quality knowledge foundation for subsequent layered fine-tuning.

[0009] In one possible implementation, the step of hierarchically fine-tuning the target base model based on the training sample pairs and the hierarchical training dataset to obtain a fine-tuned model includes: The target base model is fine-tuned by embedding domain knowledge in multiple rounds based on the general power knowledge dataset in the hierarchical training dataset, and the accuracy of the validation set after each round of domain knowledge embedding fine-tuning is collected. When the accuracy of the validation set after any round of domain knowledge embedding fine-tuning is higher than the preset validation accuracy threshold, the first basic model is obtained. Based on the fault diagnosis special dataset in the hierarchical training dataset, the first basic model is fine-tuned by focusing on fault features in multiple rounds, and the fault identification accuracy after each round of fault feature focusing fine-tuning is collected. When the fault identification accuracy after any round of fault feature focusing fine-tuning is higher than the preset identification accuracy, a second basic model is obtained. Based on the emergency repair operation dataset in the hierarchical training dataset, the second basic model is fine-tuned through multiple rounds of cross-modal fusion operation, and the recommended accuracy after each round of cross-modal fusion operation fine-tuning is collected. When the recommendation accuracy after any round of cross-modal fusion practical fine-tuning is higher than the preset recommendation accuracy, the fine-tuned model is obtained.

[0010] By adopting the above technical solutions, the power emergency repair model can be fine-tuned through a general power knowledge dataset, enabling it to master basic concepts and professional terminology in the power field and solidify its industry knowledge base. Then, by focusing on fault feature learning through a fault diagnosis-specific dataset, the power emergency repair model's ability to identify and reason about fault types and diagnostic logic can be strengthened. Finally, cross-modal fusion optimization is achieved through a repair practice dataset, allowing the power emergency repair model to adapt proficiently to the needs of practical scenarios. The hierarchical training mode helps avoid knowledge confusion, gradually deepening the professional capabilities of the power emergency repair model and thus improving the accuracy of fault diagnosis.

[0011] In one possible implementation, after acquiring the fault identification accuracy after fine-tuning the focus of each round of fault feature acquisition, the method further includes: A fault diagnosis-specific dataset is obtained by focusing on and fine-tuning the fault entities based on the fault features of each round. When the entity similarity between fault entities contained in the fault diagnosis special dataset is higher than a preset similarity threshold, the number of times the fault feature is focused and fine-tuned and the fault identification accuracy after each round of fault feature focused and fine-tuned are identified, and the accuracy convergence rate is determined based on the number of times the fault feature is focused and fine-tuned and the fault identification accuracy after each round of fault feature focused and fine-tuned. When the fault identification accuracy after any round of fault feature focusing and fine-tuning is higher than the preset identification accuracy, and the accuracy convergence rate is higher than the preset convergence rate threshold, the second basic model is determined to be generated.

[0012] By adopting the above technical solution, after collecting the fault identification accuracy after each round of fault feature focusing and fine-tuning, the fault diagnosis special dataset is updated based on the fault entities. When the similarity between fault entities in the dataset is higher than the preset similarity threshold, the accuracy convergence rate is determined by combining the number of fine-tunings and the fault identification accuracy of each round. Finally, a second base model is generated based on the dual conditions of fault identification accuracy and convergence rate. This facilitates the precise identification of model learning scenarios for highly similar faults, guides the model to focus on the core distinguishing features of easily confused faults, and thus improves the model's identification accuracy for highly similar faults. This avoids the problem of the model superficially meeting the target but failing to generalize due to a single accuracy indicator. In addition, the convergence rate is used to judge the model's learning stability, ensuring that the second base model has stably mastered the core logic of fault diagnosis, rather than relying on the surface features of the data. This helps to avoid ineffective fine-tuning iterations and improves model training efficiency.

[0013] In one possible implementation, when the number of focus adjustments for the fault features is higher than a preset number of focus adjustments, the method further includes: The fault feature discrimination vector of each highly similar fault entity in the fault diagnosis special dataset is determined based on the knowledge graph of the power field. Each highly similar fault entity in the fault diagnosis special data is combined to obtain multiple highly similar fault entity pairs, and the mean feature discrimination value is determined based on the feature discrimination value of each highly similar fault entity pair. If the mean feature discrimination score is lower than the preset discrimination score threshold, the initial focusing weight of the fault feature focusing fine-tuning corresponding fault feature attention mechanism is adjusted according to the discrimination score difference between the mean feature discrimination score and the preset discrimination score threshold to obtain the target focusing weight; Subsequent fault feature focusing fine-tuning is performed based on target focusing weights.

[0014] By adopting the above technical solution, when the number of times the fault feature focusing fine-tuning exceeds a preset value, the fault feature discrimination vector of each highly similar fault entity is determined based on the knowledge graph of the power field. The vectors are combined to form highly similar fault entity pairs and the mean feature discrimination score is calculated. Then, the initial focusing weight of the fault feature attention mechanism is adjusted according to the discrimination score difference to obtain the target focusing weight. Subsequent fine-tuning is carried out based on the target focusing weight. This facilitates the accurate positioning of the pain points in feature discrimination of highly similar faults, guides the model to focus on the core difference attributes of faults, thereby improving the model's ability to identify and distinguish highly similar faults and reducing fault misjudgment. In addition, the precise optimization of the focusing weight is achieved by quantifying the discrimination score difference, avoiding the inefficiency caused by blind fine-tuning. This facilitates the improvement of the targeting and effectiveness of fault feature focusing fine-tuning, ensuring that the model can stably improve the fault identification accuracy within a limited number of fine-tuning rounds.

[0015] In one possible implementation, after obtaining the fine-tuned model, the method further includes: Using the fine-tuned model as the teacher model, the corresponding student model is guided by the teacher model to learn features. The intermediate layer feature vectors output by the teacher model and the feature vectors output by the corresponding network layers of the student model are input into a preset distillation loss function to calculate the feature differences and obtain the total loss value of distillation training. The student model is determined based on the demand index. The contribution of each layer of the student model to the emergency repair task of power equipment is analyzed. The importance score of each layer parameter of the student model is calculated based on the summation of the absolute values ​​of the parameter gradients. The network structure of the student model is pruned using a preset iterative pruning strategy. Adaptive quantization is performed on the pruned student model to obtain a quantized optimized model. The quantized optimized model is then deployed and adapted to complete the dynamic compression and full-process optimization of the fine-tuned model.

[0016] By adopting the above technical solution, the fine-tuned model is used as the teacher model to guide the student model in feature learning. The feature vectors of the corresponding network layers of the teacher and student models are input into a preset distillation loss function to calculate the total loss value. Then, the contribution of each layer of the student model to the emergency repair task of power equipment is analyzed, and a preset iterative pruning strategy is executed based on the parameter importance score. Finally, the pruned student model is subjected to adaptive quantization and deployment adaptation optimization, completing the dynamic compression and quantization optimization of the fine-tuned model. This facilitates a significant reduction in the number of model parameters and computational complexity while inheriting the professional semantic understanding ability of the teacher model, thus making it easier to achieve model lightweighting. In addition, the distillation loss precisely constrains the learning effect of the student model, and the iterative pruning preserves the core network structure. The adaptive quantization balances the accuracy and the degree of lightweighting, which helps to avoid the accuracy loss caused by traditional compression methods.

[0017] In one possible implementation, the method further includes: Collect emergency repair application data of the fine-tuned model in actual power equipment emergency repair scenarios, including fault diagnosis error cases and feedback optimization suggestions; The emergency repair application data is added to the sample multimodal data to obtain updated multimodal data, and the power equipment emergency repair knowledge graph is optimized based on the updated multimodal data.

[0018] By adopting the above technical solution, the fault diagnosis error cases and feedback optimization suggestions of the fine-tuned model in actual power equipment emergency repair scenarios are collected and supplemented into the sample multimodal data to obtain updated multimodal data. Based on the updated multimodal data, the knowledge graph of power equipment emergency repair is optimized, which can enrich the scenario coverage and authenticity of the model training data, make up for the special scenarios and edge cases of actual emergency repair that were not covered by the original training data, and thus improve the model's adaptability to complex and diverse actual emergency repair scenarios.

[0019] Secondly, this application provides a fine-tuning system, which adopts the following technical solution: A fine-tuning system comprising: At least one processor; Memory; At least one application, wherein the at least one application is stored in memory and configured to be executed by at least one processor, the at least one application being configured to: execute the above-described knowledge graph-based power emergency repair model fine-tuning method.

[0020] Thirdly, this application provides a computer-readable storage medium, which adopts the following technical solution: A computer-readable storage medium includes: a computer program that can be loaded by a processor and executed by the above-described knowledge graph-based power emergency repair model fine-tuning method.

[0021] Fourthly, this application provides a computer program product, which adopts the following technical solution: A computer program product includes a computer program that, when executed by a processor, implements the aforementioned knowledge graph-based power emergency repair model fine-tuning method.

[0022] In summary, this application includes at least one of the following beneficial technical effects: By constructing a knowledge graph for power equipment emergency repair, structured knowledge such as equipment parameters, fault types, and repair procedures is integrated into fine-tuning, which facilitates the improvement of the fault diagnosis accuracy of the power emergency repair model. At the same time, hierarchical training is carried out using general power knowledge, fault diagnosis-specific knowledge, and emergency repair practice, which helps guide the power emergency repair model to gradually master professional capabilities and avoid knowledge confusion. This facilitates the improvement of the ability to distinguish highly similar faults and the reasoning ability in complex emergency repair scenarios. By fusing the triplet data in the power equipment emergency repair knowledge graph with effective multimodal data of management and control to form training sample pairs, and performing hierarchical fine-tuning of the target basic model based on the training sample pairs and hierarchical training datasets, the power emergency repair model can absorb structured knowledge and multi-source data features at the same time. This facilitates the improvement of the accuracy of the power emergency repair model in processing multimodal data, achieving accurate fault diagnosis, thereby improving the efficiency of operation, maintenance, and emergency repair work, and ultimately improving the reliability of power supply. After collecting fault identification accuracy data for each round of feature-focused fine-tuning, the fault diagnosis dataset is updated based on fault entities. When the similarity between fault entities in the dataset exceeds a preset similarity threshold, the accuracy convergence rate is determined by combining the number of fine-tuning iterations with the fault identification accuracy for each round. Finally, a second base model is generated based on both the fault identification accuracy and convergence rate targets. This facilitates the precise identification of highly similar faults in the model learning scenario, guiding the model to focus on the core distinguishing features of easily confused faults. This improves the model's accuracy in identifying highly similar faults and avoids the problem of superficially meeting the target accuracy but failing to generalize due to a single accuracy metric. Furthermore, the convergence rate is used to judge the model's learning stability, ensuring that the second base model has stably mastered the core logic of fault diagnosis rather than relying on surface features of the data. This helps avoid ineffective fine-tuning iterations and improves model training efficiency. Attached Figure Description

[0023] Figure 1 This is a flowchart illustrating a knowledge graph-based method for fine-tuning a power emergency repair model in an embodiment of this application. Figure 2 This is a schematic diagram of a fault feature focusing fine-tuning process in an embodiment of this application; Figure 3 This is a schematic diagram of the structure of a fine-tuning system in an embodiment of this application. Detailed Implementation

[0024] The following is in conjunction with the appendix Figures 1 to 3 This application will be described in further detail.

[0025] After reading this specification, those skilled in the art may make modifications to this embodiment without contributing any inventive step, but such modifications are protected by patent law as long as they fall within the scope of the claims of this application.

[0026] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0027] It should be noted that, in the optional embodiments of this application, the data related to object information, when applied to specific products or technologies, requires the permission or consent of the object. Furthermore, the collection, use, and processing of this data must comply with the relevant laws, regulations, and standards of the relevant countries and regions. In other words, if the embodiments of this application involve data related to an object, it must be obtained with the object's authorization and consent, the authorization and consent of relevant departments, and in accordance with the relevant laws, regulations, and standards of the country and region. If the embodiments involve personal information, the acquisition of all personal information requires the individual's consent. If sensitive information is involved, the separate consent of the information subject is required. The embodiments also need to be implemented with the object's authorization and consent.

[0028] Specifically, this application provides a knowledge graph-based method for fine-tuning a power emergency repair model, executed by a fine-tuning system. This system can be a server or a terminal device. The server 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. The terminal device can be a smartphone, tablet, laptop, desktop computer, etc., but is not limited to these. The terminal device and the server can be directly or indirectly connected via wired or wireless communication, and this application does not impose any limitations on this connection.

[0029] refer to Figure 1 , Figure 1 This is a flowchart illustrating a knowledge graph-based power emergency repair model fine-tuning method according to an embodiment of this application. The method includes steps S110-S170, wherein: Step S110: Collect sample multimodal data in the field of power grid operation, maintenance and emergency repair, and classify and label the sample multimodal data to obtain the classified and labeled sample multimodal data to be processed. The sample multimodal data includes text data, image data and structured data.

[0030] Specifically, power equipment fault diagnosis and emergency repair require the integration of multi-dimensional information such as textual descriptions, visual features, and equipment parameters. Collecting and analyzing multimodal sample data can comprehensively reconstruct the fault scenario and avoid the one-sidedness of information caused by a single data type. Among them, multimodal sample data includes text data, image data, and structured data. Different data types may require different collection or acquisition methods. For example, text data containing equipment operation and maintenance records, fault work orders, maintenance standards, emergency repair process specifications, and professional terminology dictionaries can be exported from the power grid operation and maintenance emergency repair system. Image data containing photos of equipment defects, inspection images, and fault site video frames can be collected through inspection drones, on-site mobile phones, and monitoring cameras. Image data should focus on covering fault scenarios of core equipment such as transformers, circuit breakers, and insulators. Structured data such as equipment parameter tables (such as rated voltage and capacity), historical fault statistics, and geographic information data can be extracted from relevant power management platforms.

[0031] The process involves classifying and labeling the multimodal data samples. This means adding standardized labels to different multimodal data samples based on data type and business scenario. This gives the multimodal data samples clear classification attributes and business meanings, facilitating subsequent model recognition and learning. Different data types require different classification and labeling methods: text data can be classified and labeled using a labeling system of "fault description - fault type - repair plan - safety procedures"; image data can be classified and labeled using a labeling system of "equipment type - defect location - defect level"; and structured data can be classified and labeled using a labeling system of "equipment number - parameter type - operating status - historical fault association". After classifying and labeling all the multimodal data samples, the multimodal data samples to be processed are obtained.

[0032] Step S120: Perform categorized data management and control processing on the sample multimodal data to be processed, obtain effective multimodal data for management and control, and construct a knowledge graph for emergency repair of power equipment based on the effective multimodal data for management and control.

[0033] Specifically, the process involves categorizing and managing multimodal data to be processed. This means employing differentiated quality optimization strategies for different types of multimodal data, such as text, images, and structured data. Operations include redundancy removal, error correction, missing data filling, and anomaly filtering to improve data quality and obtain effective multimodal data that meets the requirements for model training. The data management methods differ depending on the data type of the multimodal data being processed. Regarding multimodal text sample data to be processed: regular expressions can be used to remove redundant characters and garbled text, natural language processing techniques can be used for word segmentation, part-of-speech tagging and semantic error correction, a general large model can be used to score the quality of text data, and manual review can be used to remove duplicate, erroneous and incomplete data, and texts with abnormal length (too long or too short) can be truncated or supplemented. Regarding multimodal data to be processed in image samples: preprocessing operations such as image correction, noise reduction, and enhancement (contrast adjustment, edge detection) can be used. The YOLOv7-Paddle visual recognition framework can be used to perform preliminary defect detection on the images and filter out image data without effective defect information. Regarding structured sample multimodal data to be processed: missing value imputation (based on the statistical mean or median of parameters of similar devices) and outlier removal (identified and processed through the 3σ principle) can be used to ensure data integrity and accuracy.

[0034] The knowledge graph for power equipment emergency repair is a structured and systematic professional knowledge network that integrates core information such as equipment, faults, repair resources, and processes in the field of power equipment emergency repair, using a pre-set graph database as its storage carrier and "entity-attribute-relationship" triples as its core structure. The specific process of constructing the knowledge graph for power equipment emergency repair based on effective multimodal data management can include: Core entities in the field of power equipment emergency repair are extracted from the effective data of multimodal management and control. These core entities include equipment entities, fault entities, emergency repair resource entities, and process entities. The relationships between these core entities are determined based on the business logic of the power equipment emergency repair field. A knowledge graph storage architecture is built based on a pre-set graph database, and a triple structure with entity-attribute-relationship as the core is constructed to form a power equipment emergency repair knowledge system. The association mapping relationship between the power equipment emergency repair knowledge system and the effective data of multimodal management and control is established to obtain the power equipment emergency repair knowledge graph.

[0035] Specifically, core entities in the field of power equipment emergency repair can be extracted from multimodal effective data managed by a preset feature recognition algorithm. These core entities include equipment entities, fault entities, repair resource entities, and process entities. Equipment entities cover power equipment such as transformers, circuit breakers, and insulators; fault entities cover equipment fault types such as short-circuit faults, insulation damage, and equipment overheating; repair resource entities cover repair tools, spare parts, and repair personnel; and process entities cover operational processes such as fault diagnosis and repair procedures. The relationships between these core entities can be determined based on the business logic of the power equipment emergency repair field. This business logic primarily originates from industry standards, practical data, and professional experience. The business logic defines the essential association rules between the core entities, serving as the underlying logical support for these relationships. The relationships include, but are not limited to: equipment-fault association, fault-repair plan association, equipment-parameter association, and process-step association. For example, transformer-winding short circuit is an equipment-fault association, insulation damage-insulator replacement is a fault-repair plan association, circuit breaker-rated voltage is an equipment-parameter association, and fault investigation-visual inspection-parameter detection is a process-step association.

[0036] The default graph database can be Neo4j, which leverages its efficient node-relationship storage and association query capabilities to adapt to the complex network-like relationships of power emergency repair knowledge. It supports efficient association queries and dynamic updates, meeting the complex association query needs of power emergency repair knowledge. The storage architecture includes an entity node layer, an attribute description layer, and a relationship connection layer. The entity node layer primarily stores four types of core nodes: equipment entities, fault entities, emergency repair resource entities, and process entities. Each node is bound to a unique identifier (such as equipment ID, fault code) and core attributes (such as equipment model, fault level). The attribute description layer mainly uses key-value pairs to mount detailed attributes of each entity, such as "rated voltage" and "installation location" for equipment entities, and "fault characteristics" and "affected area" for fault entities, supporting dynamic attribute expansion. The relationship connection layer represents the relationships between entities through directed edges. Each edge is labeled with the relationship type (such as "occurred," "required," "included") and business attributes (such as relationship activation conditions, priority), forming a "node-edge-node" association structure.

[0037] The triplet format uniformly adopts two standard formats: "Entity 1-Relationship-Entity 2" and "Entity-Attribute-Attribute Value," ensuring the structured and consistent representation of knowledge. The triplet generation process involves converting extracted entities, attributes, and relationships into standardized triples using data mapping rules. For example, from "Transformer T1001 experienced a winding short circuit, requiring a multimeter for testing," two triples are generated: [T1001]-[Occurred]-[F003] and [F003]-[Required Tool]-[Multimeter R008]. The batch data entry process includes... The process involves batch importing triplet data using Neo4j's Cypher statements, while simultaneously performing data validation. This validation checks for "isolated nodes (no relationships)," "invalid relationships (not conforming to emergency repair business logic)," and "missing attributes (missing key equipment parameters or fault characteristics)." Based on the business scenarios of power emergency repair (substation repair, transmission line repair, and distribution line repair), the triplet structure is hierarchically divided, and indexes are created for the key attributes of core entities (equipment name, fault name, and process name) to improve the efficiency of subsequent knowledge graph retrieval and association. All triplet structures that pass validation, have completed hierarchical division, and index construction collectively constitute the knowledge system for the power equipment emergency repair domain.

[0038] The mapping relationship between the knowledge system for emergency repair of power equipment and the effective multimodal data for management and control includes text-based data mapping, image-based data mapping, and structured data mapping. Among them, text-based data mapping associates fault work order texts with equipment entities in the knowledge graph through "equipment number" and fault entities through "fault description keywords". Image-based data mapping associates equipment defect images with equipment entities through "equipment type label" and fault entities through "defect location / level". Structured data mapping: the equipment parameter table associates equipment entity attributes one-to-one through "equipment number", the historical fault statistics table associates fault entities through "fault code", and the tool inventory table associates emergency repair resource entities through "resource number".

[0039] Finally, by integrating the aforementioned core entities, the relationships between them, and the mapping relationships between the power equipment emergency repair knowledge system and the effective multimodal data for management and control, a complete power equipment emergency repair knowledge graph is formed. This knowledge graph deeply binds multimodal data with professional knowledge, facilitating the model's rapid association of key information such as equipment parameters, fault characteristics, and repair procedures. This avoids knowledge fragmentation and enhances the model's understanding of professional power emergency repair scenarios, providing a high-quality knowledge foundation for subsequent layered fine-tuning.

[0040] Step S130: Identify the demand indicators corresponding to the effective data of multimodal control, and determine the candidate models and the fit evaluation system based on the demand indicators.

[0041] Specifically, the required metrics include the model's requirements for structural capabilities, inference efficiency, deployability, and industry knowledge coverage. The system can identify data type features from the effective multimodal data under control based on the preset feature recognition algorithm, and determine the model’s structural capability requirements based on the data type features. Since the effective multimodal data under control includes text, image, and structured multimodal data, and each type of data has professional annotation information in the field of power equipment emergency repair, the multimodal data fusion processing capability required by the model can be obtained. Based on the preset feature recognition algorithm, application scenario adaptation features are identified from the effective data of multimodal management and control. Based on the scenario adaptation features, the model’s requirements for inference efficiency are determined. Since the core application scenario of effective data of multimodal management and control is emergency repair of power equipment, this scenario requires the rapid completion of tasks such as fault identification and emergency repair plan recommendation based on data. Therefore, the high inference efficiency required by the model can be obtained. Based on the preset feature recognition algorithm, industrial-grade usage characteristics are identified from the analysis and control of multimodal effective data. Based on the industrial-grade usage characteristics, the model's deployment requirements are determined. Since the control of multimodal effective data belongs to the core operation data of the power grid, it must be used in a private environment in accordance with data security management requirements. Furthermore, the storage and retrieval of the data must match the hardware resource conditions of the power grid site. Therefore, the resource carrying capacity required by the model can be obtained. Based on the preset feature recognition algorithm, professional attribute features are identified from the effective data of multimodal analysis and control. Based on the professional attribute features, the industry knowledge coverage required by the model is determined. Since the effective data of multimodal analysis and control are all professional data in the field of power grid operation, maintenance and emergency repair, including exclusive knowledge in fields such as equipment parameters, fault types and emergency repair procedures, the industry knowledge coverage required by the model can be obtained.

[0042] Based on demand metrics, mainstream domestic models can be selected as candidate foundational models. Selection criteria include parameter size, multimodal processing capability, inference efficiency, open-source compatibility, and private deployment support. Furthermore, based on these demand metrics, a compatibility evaluation system is constructed from dimensions such as structural capabilities, industry knowledge coverage, inference efficiency, and deployability. Structural capabilities include the number of model layers, the number of attention heads, and the multimodal fusion architecture; industry knowledge coverage is assessed through a test score using a set of 600 power industry professional questions and answers; inference efficiency is measured by the model's response latency and throughput in power equipment fault diagnosis tasks; and deployability includes model size, hardware resource requirements, and compatible system types.

[0043] Step S140: Based on the fitness evaluation system, perform standardized testing on each candidate model to obtain the comprehensive fitness score of each candidate model, and take the candidate model with the highest comprehensive fitness score as the target base model.

[0044] Specifically, standardized testing for each candidate model includes: Unified test dataset: High-quality multimodal effective data obtained through data preprocessing and knowledge graph construction steps can be selected to build a unified power equipment emergency repair test dataset. The test dataset covers preprocessed text data, image data, and structured data. Unified testing tasks: For core tasks such as fault type identification, emergency repair plan recommendation, and professional knowledge Q&A, unified testing standards (such as the method of calculating the accuracy of fault identification and the test environment for response latency) are implemented. Unified quantification of indicators: The indicators of each dimension are converted into standardized scores of 0-100 (e.g., industry knowledge coverage is calculated as "number of correct answers / total number of questions × 100", and reasoning efficiency is normalized as "baseline delay / actual delay × 100"), so as to make the scores of different dimensions comparable.

[0045] Finally, a weighted summation is used to calculate the overall score, with the core principle being "giving higher weight to more important scenario requirements": Weight allocation logic: Industry knowledge coverage has the highest weight (0.3) because "professional accuracy" is the core of the power emergency repair model; structural capability and reasoning efficiency have the next highest weights (0.25 each) to balance the model's processing potential and real-time performance; deployability has a weight of 0.2 to ensure that the model can be deployed and used. Calculation logic: Overall fitness score = structural capability score × 0.25 + industry knowledge coverage score × 0.3 + reasoning efficiency score × 0.25 + deployability score × 0.2. The higher the overall fitness score, the stronger the overall fitness of the candidate model. The candidate model with the highest overall fitness score is used as the target base model of this application.

[0046] Step S150: Construct a hierarchical training dataset based on the knowledge graph of power equipment emergency repair. The hierarchical training dataset includes a general power knowledge dataset, a fault diagnosis-specific dataset, and an emergency repair practice dataset.

[0047] Step S160: Identify the triplet data in the knowledge graph of power equipment emergency repair, and perform fusion processing based on the triplet data and effective multimodal data of management and control to form training sample pairs.

[0048] Specifically, "equipment-attribute-attribute value" and "term-definition" triples can be extracted from the knowledge graph of power equipment emergency repair to construct a general power knowledge dataset. The general power knowledge dataset contains power professional terms and basic equipment information. The "equipment-fault-feature", "fault-cause-level", and "fault-discriminative feature" triples can be extracted from the knowledge graph of power equipment emergency repair to construct a fault diagnosis dataset. The fault diagnosis dataset contains equipment fault features and fault diagnosis logic. The "fault-repair process-step", "fault-required resources-specifications", and "process-operation standard-requirements" triples can be extracted from the power equipment emergency repair knowledge graph to construct an emergency repair practice dataset. The emergency repair practice dataset includes emergency repair processes, emergency repair operation specifications, and emergency repair case data.

[0049] Based on a preset feature recognition algorithm, the triplet data contained in the knowledge graph of power equipment emergency repair are identified, and the identified triplet data is converted into a text sequence format that can be recognized by the target basic model (such as "[Equipment: Transformer]-[Fault: Winding Short Circuit]-[Emergency Repair Plan: Power Outage Inspection-Winding Replacement-Insulation Test]"). The triplet data is then fused with the effective data of multimodal management and control to form training sample pairs.

[0050] Step S170: Based on the training sample pairs and the hierarchical training dataset, perform hierarchical fine-tuning on the target base model to obtain the fine-tuned model.

[0051] Specifically, the training sample pair is used throughout the entire process of hierarchical fine-tuning of the target base model. In the entire fine-tuning process, the training sample pair is used as the core, and then combined with the fine-tuning results of each fine-tuning stage for fusion input, providing a data foundation for parameter optimization and capability training of the model at each stage.

[0052] Furthermore, to improve the accuracy of fault diagnosis, the target base model is fine-tuned hierarchically based on training sample pairs and hierarchical training datasets. The specific process of obtaining the fine-tuned model may include: The target basic model is fine-tuned through multiple rounds of domain knowledge embedding based on the general power knowledge dataset in the hierarchical training dataset, and the validation set accuracy is collected after each round of domain knowledge embedding fine-tuning. When the validation set accuracy after any round of domain knowledge embedding fine-tuning is higher than the preset validation accuracy threshold, the first basic model is obtained. The first basic model is then fine-tuned through multiple rounds of fault feature focusing based on the fault diagnosis special dataset in the hierarchical training dataset, and the fault identification accuracy is collected after each round of fault feature focusing fine-tuning. When the fault identification accuracy after any round of fault feature focusing fine-tuning is higher than the preset identification accuracy, the second basic model is obtained. The second basic model is then fine-tuned through multiple rounds of cross-modal fusion practical operation based on the emergency repair practical dataset in the hierarchical training dataset, and the recommendation accuracy is collected after each round of cross-modal fusion practical operation fine-tuning. When the recommendation accuracy after any round of cross-modal fusion practical operation fine-tuning is higher than the preset recommendation accuracy, the fine-tuned model is obtained.

[0053] Specifically, a general power knowledge dataset can be extracted from the hierarchical training dataset based on a preset feature recognition algorithm. Then, the target basic model can be fine-tuned by embedding domain knowledge in multiple rounds based on the general power knowledge dataset. The entity embedding and text embedding in the power equipment emergency repair knowledge graph are fused and input into the target basic model to optimize the target basic model's understanding of basic concepts in the power field. Generally, the fine-tuning rounds are 3-5 rounds. The accuracy of the validation set is recorded after each round of domain knowledge embedding fine-tuning. Only when the accuracy of the validation set after any round of domain knowledge embedding fine-tuning is higher than the preset validation accuracy threshold can it be said that the target basic model has completed the domain knowledge embedding fine-tuning and has been optimized into the first basic model.

[0054] Based on a preset feature recognition algorithm, a fault diagnosis-specific dataset is extracted from the hierarchical training dataset. Then, based on the fault diagnosis-specific dataset, the first basic model is fine-tuned with multiple rounds of fault feature focusing. A fault feature attention mechanism is introduced, which makes the first basic model focus on the text description, image features and knowledge graph associations related to equipment faults. Generally, the fine-tuning rounds are 4-6 rounds. After each round of fault feature focusing fine-tuning, the fault recognition accuracy is recorded. Only when the fault recognition accuracy after any round of fault feature focusing fine-tuning is higher than the preset recognition accuracy, it indicates that the first basic model has completed the fault feature focusing fine-tuning and has been optimized into the second basic model.

[0055] Based on a preset feature recognition algorithm, a practical emergency repair dataset is extracted from the hierarchical training dataset. Then, based on this dataset, the second basic model undergoes multiple rounds of cross-modal fusion practical fine-tuning. A multi-modal data fusion module is integrated, aligning and fusing image features (extracted via CNN), structured data features (encoded via fully connected layers), text features, and knowledge graph features across modalities. This optimizes the second model's emergency repair scheme generation and reasoning capabilities. Generally, 3-6 rounds of fine-tuning are performed, and the recommendation accuracy is recorded in each round. Only when the recommendation accuracy after any round of cross-modal fusion practical fine-tuning exceeds the preset recommendation accuracy is the second basic model considered to have completed cross-modal fusion practical fine-tuning and become the final fine-tuned model. The specific values ​​of the preset recognition accuracy, preset recommendation accuracy, and preset recognition accuracy are not specifically limited in this embodiment and can be determined by relevant personnel according to actual needs and uploaded to the fine-tuning system. The hierarchical training mode helps avoid knowledge confusion, gradually deepening the professional capabilities of the power emergency repair model, thereby improving fault diagnosis accuracy.

[0056] In this embodiment of the application, by constructing a knowledge graph for power equipment emergency repair, structured knowledge such as equipment parameters, fault types, and emergency repair procedures is integrated into fine-tuning, which facilitates the improvement of the fault diagnosis accuracy of the power emergency repair model. At the same time, hierarchical training is carried out using general power knowledge, fault diagnosis-specific knowledge, and emergency repair practice, which helps guide the power emergency repair model to gradually master professional capabilities and avoid knowledge confusion. This facilitates the improvement of the ability to distinguish highly similar faults and the reasoning ability in complex emergency repair scenarios. By fusing the triplet data in the power equipment emergency repair knowledge graph with effective multimodal data of management and control to form training sample pairs, and performing hierarchical fine-tuning of the target basic model based on the training sample pairs and hierarchical training datasets, the power emergency repair model can simultaneously absorb structured knowledge and multi-source data features, thereby improving the accuracy of the power emergency repair model in processing multimodal data, achieving accurate fault diagnosis, thereby improving the efficiency of operation and maintenance emergency repair work, and ultimately improving the reliability of power supply.

[0057] Furthermore, to avoid ineffective fine-tuning iterations, after collecting the fault feature focus and fine-tuning accuracy of each round of fault identification, the process may further include steps S210-S230, such as... Figure 2 As shown, where: Step S210: Obtain a fault diagnosis-specific dataset based on the fault entities fine-tuned by focusing on fault features in each round.

[0058] Specifically, since there are a large number of highly similar entities in power equipment faults (such as “winding short circuit” and “lead short circuit”, “insulation damage” and “insulation aging”), the feature differences of these entities are small. Therefore, during the model fine-tuning process, the problem of “acceptance accuracy meets the standard but generalization ability is poor” is likely to occur. Therefore, it is possible to record and statistically analyze the fault entities after each round of fault feature focusing fine-tuning to obtain a fault diagnosis special dataset.

[0059] Step S220: When the entity similarity between fault entities in the fault diagnosis special dataset is higher than the preset similarity threshold, identify the number of times the fault feature is fine-tuned and the fault identification accuracy after each round of fault feature fine-tuning, and determine the accuracy convergence rate based on the number of times the fault feature is fine-tuned and the fault identification accuracy after each round of fault feature fine-tuning.

[0060] Specifically, the similarity of fault entities in the fault diagnosis dataset can be compared to obtain the entity similarity between fault entities. When the entity similarity between any pair of fault entities is higher than the preset similarity threshold, the accuracy convergence rate corresponding to the entire fault feature focusing fine-tuning process can be analyzed. The accuracy convergence rate is calculated as follows: (Current round fault identification accuracy - Initial round fault identification accuracy) / Number of fault feature focusing fine-tuning operations.

[0061] Step S230: When the fault identification accuracy after any round of fault feature focusing and fine-tuning is higher than the preset identification accuracy and the accuracy convergence rate is higher than the preset convergence rate threshold, the second basic model is determined to be generated.

[0062] Specifically, when the entity similarity between fault entities in the fault diagnosis dataset is higher than a preset similarity threshold, if we still rely solely on the fault recognition accuracy being higher than the preset recognition accuracy to determine whether the first basic model has completed fault feature focusing fine-tuning, it may not be possible to determine whether the first basic model has truly mastered the fault differentiation logic or is merely achieving a false result due to "rote memorization" of training data. Therefore, when there are many highly similar fault entities in a short period of time, it is necessary to add a judgment index for model fine-tuning. That is, only when the fault recognition accuracy after any round of fault feature focusing fine-tuning is higher than the preset recognition accuracy, and the accuracy convergence rate is higher than the preset convergence rate threshold, can it be characterized that the first basic model has completed fault feature focusing fine-tuning. In other words, only then can it be determined that the first basic model has been optimized into the second basic model.

[0063] In this embodiment, after collecting the fault identification accuracy after each round of fault feature focusing and fine-tuning, the fault diagnosis special dataset is updated based on the fault entities. When the similarity between fault entities in the dataset is higher than a preset similarity threshold, the accuracy convergence rate is determined by combining the number of fine-tunings and the fault identification accuracy of each round. Finally, a second basic model is generated based on the dual conditions of fault identification accuracy and convergence rate. This facilitates the precise identification of model learning scenarios for highly similar faults, guides the model to focus on the core distinguishing features of easily confused faults, thereby improving the model's identification accuracy for highly similar faults and avoiding the problem of the model superficially meeting the target but failing to generalize due to a single accuracy indicator. In addition, the model learning stability is judged by the convergence rate to ensure that the second basic model has stably mastered the core logic of fault diagnosis, rather than relying on the surface features of the data. This helps to avoid ineffective fine-tuning iterations and improve the model training efficiency.

[0064] Furthermore, when the number of fault feature focusing fine-tuning operations exceeds the preset number of focusing fine-tuning operations, the method provided in this application embodiment further includes: Based on the knowledge graph of the power field, the fault feature discrimination vector of each highly similar fault entity in the fault diagnosis special data is determined; each highly similar fault entity in the fault diagnosis special data is combined to obtain multiple highly similar fault entity pairs, and the mean feature discrimination is determined based on the feature discrimination of each highly similar fault entity pair; if the mean feature discrimination is lower than the preset discrimination threshold, the initial focusing weight of the fault feature focusing fine-tuning corresponding to the fault feature attention mechanism is adjusted according to the discrimination difference between the mean feature discrimination and the preset discrimination threshold to obtain the target focusing weight; subsequent fault feature focusing fine-tuning is performed based on the target focusing weight.

[0065] Specifically, after multiple rounds of fault feature focusing and fine-tuning, the first basic model still has not completed the fault feature focusing and fine-tuning. At this point, the core attributes of each highly similar fault entity in the fault diagnosis special dataset can be identified based on the power field knowledge graph (core attributes include, but are not limited to, fault feature description, associated equipment parameters, triggering conditions, and differences in emergency repair plans). Then, a fault feature discrimination vector corresponding to each highly similar fault entity is constructed based on the core attributes. All highly similar fault entities in the fault diagnosis special dataset are combined into multiple highly similar fault entity pairs. Then, based on the fault feature discrimination vectors corresponding to the highly similar fault entities in the highly similar fault entity pairs, the feature discrimination degree of each highly similar fault entity pair is calculated. Feature discrimination degree = 1 – attribute cosine similarity of the highly similar fault entity pair. Finally, the mean of the feature discrimination degrees of all highly similar fault entity pairs is calculated to obtain the mean feature discrimination degree.

[0066] When the mean feature discrimination score is lower than the preset discrimination score threshold, the discrimination score difference between the mean feature discrimination score and the preset discrimination score threshold can be calculated. Based on the preset weight adjustment value mapping relationship, the corresponding weight adjustment value is determined. Finally, based on this weight adjustment value, the initial focusing weights of the fault feature attention mechanism corresponding to the fault feature focusing fine-tuning are adjusted to obtain the target focusing weights. Subsequent fault feature focusing fine-tuning is then performed based on the target focusing weights. This avoids blind iteration after exceeding the limit of fine-tuning times, shortens ineffective training time through precise weight adjustment, balances model accuracy and training efficiency, and facilitates improved targeting and effectiveness of fault feature focusing fine-tuning, ensuring that the model steadily improves fault identification accuracy within a limited number of fine-tuning rounds. The preset weight adjustment value mapping relationship is the correspondence between the discrimination score difference and the weight adjustment value; the larger the discrimination score difference, the larger the corresponding weight adjustment value. The specific content of this mapping relationship is not specifically limited in this embodiment and can be determined by relevant personnel based on historical experimental data and uploaded to the fine-tuning system.

[0067] Furthermore, after obtaining the fine-tuned model, the method provided in this application embodiment may further include: Using the fine-tuned model as the teacher model, the corresponding student model is guided by the teacher model to learn features. The intermediate layer feature vectors output by the teacher model and the feature vectors output by the corresponding network layers of the student model are input into a preset distillation loss function to calculate the feature differences and obtain the total loss value of distillation training. The student model is determined based on the demand index. The contribution of each layer of the student model to the emergency repair task of power equipment is analyzed. The importance score of each layer parameter of the student model is calculated by summing the absolute values ​​of the parameter gradients. A preset iterative pruning strategy is used to prune the network structure of the student model. Adaptive quantization processing is performed on the pruned student model to obtain the quantized optimized model. The quantized optimized model is then deployed and adapted to complete the dynamic compression and full-process optimization of the fine-tuned model.

[0068] Specifically, a fine-tuned model is used as the teacher model, while a lightweight model with 4B-8B parameters (such as Qwen3-4B) can be selected as the student model based on the required metrics. A strategy combining intermediate layer distillation and task-oriented distillation can be adopted. The teacher model outputs intermediate layer feature vectors and task prediction results to guide the student model's learning. The preset distillation loss function can be a weighted sum of KL divergence loss (weight 0.3-0.5) and mean squared error loss (weight 0.5-0.7). The distillation training data can use an augmented dataset from the power equipment emergency repair domain. The training rounds are generally 8-12 rounds. The total loss value of the distillation training for each round is recorded. The total loss value of each round is transmitted to the network parameters of the student model through a backpropagation mechanism, driving the student model to adjust its own parameters and continuously narrow the difference with the teacher model in intermediate layer feature extraction and task prediction results. Ultimately, the student model effectively inherits the teacher model's professional semantic understanding, fault diagnosis, and solution reasoning capabilities in the power equipment emergency repair domain.

[0069] First, the contribution of each layer of the student model to the emergency repair task of power equipment can be analyzed, and the importance score of each layer parameter can be calculated. Then, based on the importance score of each layer parameter and combined with an iterative pruning strategy, the student model can be precisely pruned. The importance score of each layer parameter provides a clear selection criterion for each round of pruning (e.g., only pruning layers with a score below 0.03). Without the importance score of each layer parameter, the pruning ratio may lose its appropriate basis, leading to situations such as "over-pruning in a certain round" or "continuous pruning of core layers," causing the model to be unable to recover its performance through subsequent fine-tuning and compromising the controllability of the pruning process. When determining the importance score of each layer parameter, the sum of the absolute gradient values ​​of all layers in the student model can be calculated to obtain the global absolute gradient value sum. Then, the ratio of the absolute gradient value sum of each layer to the global absolute gradient value sum is used as the importance score of that layer. The specific method is not specifically limited in this embodiment. The iterative pruning strategy aims to "precisely remove redundancy and retain the core structure." Based on the importance scores of parameters in each layer, the pruning ratio is gradually reduced in multiple rounds to avoid model performance collapse caused by large-scale pruning in a single round. At the same time, the model accuracy is stabilized through fine-tuning after each round of pruning. For example, high-contribution layers with an importance score ≥ 0.08 (such as fault feature attention layers and multimodal fusion layers) are directly excluded from the pruning range, and only low-contribution layers with an importance score < 0.03 are included in the first round of pruning candidate pool. In the first round of pruning, the pruning ratio is set to 15%, and the candidate pool is pruned. Redundant convolutional kernels are removed from convolutional layers, and low-weight attention heads are pruned from attention layers to ensure the integrity of the core network structure. In the second round of pruning, the pruning ratio is reduced to 10%, the importance scores of each layer of the model are recalculated after pruning, the candidate pool is updated, and pruning is only performed on newly added low-contribution layers (score < 0.03) to avoid repeatedly pruning the same level.

[0070] The pruned student model is quantized with low bit depth, the weight parameters are quantized with INT8, the activation values ​​are quantized with FP16, and the parameters of the key channels for power equipment fault characteristics are trained with quantization-aware training (QAT) to retain high-precision representation. A dynamic quantization factor is introduced to adaptively adjust the quantization precision according to the input data type (text, image, structured data). Standard quantization is used for text data, while high-precision quantization mode is used for image and equipment parameter data.

[0071] Finally, the model inference engine can be optimized for the hardware environment (mobile terminals, edge computing devices) at power grid operation and maintenance sites. A GPU / CPU hybrid scheduling and caching mechanism is adopted to cache the inference results of frequently accessed fault diagnosis and emergency repair plan recommendation modules. Offline model operation mode is supported, and the quantized optimized model and core knowledge graph fragments are packaged and deployed to ensure that auxiliary services can still be provided normally in the absence of network environment.

[0072] By optimizing the entire process of dynamic compression and quantization of the fine-tuning model, the number of parameters and computational complexity of the fine-tuning model can be significantly reduced while inheriting the professional semantic understanding capabilities of the teacher's model. This facilitates the lightweighting of the model. In addition, the learning effect of the student model is precisely constrained by distillation loss. Combined with iterative pruning to preserve the core network structure and adaptive quantization to balance accuracy and lightweighting, the accuracy loss caused by traditional compression methods can be avoided.

[0073] Furthermore, to enhance the model's adaptability to complex and diverse real-world emergency repair scenarios, the method provided in this application embodiment also includes: We collect emergency repair application data of the fine-tuning model in actual power equipment emergency repair scenarios. The emergency repair application data includes fault diagnosis error cases and feedback optimization suggestions. We supplement the sample multimodal data with the emergency repair application data to obtain updated multimodal data, and optimize the power equipment emergency repair knowledge graph based on the updated multimodal data.

[0074] Specifically, online feedback portals can be established, offline work records can be connected, and expert reviews can be organized to collect data on the application of the fine-tuned model in actual power equipment emergency repair scenarios. Emergency repair application data refers to data related to model performance generated in actual power equipment emergency repair scenarios after the fine-tuned model is deployed. The core data includes fault diagnosis error cases (scenario data where the model diagnosis results do not match the actual situation) and feedback optimization suggestions (model improvement opinions put forward by front-line personnel and experts).

[0075] When supplementing emergency repair application data into the sample multimodal data, this includes entity addition, relationship correction, and attribute improvement. The power equipment emergency repair knowledge graph is optimized based on the updated multimodal data; that is, the power equipment emergency repair knowledge graph is reconstructed based on the updated multimodal data. The specific process can be referred to in the above embodiment for constructing the power equipment emergency repair knowledge graph based on effective multimodal data of management and control, and will not be elaborated here. By optimizing the power equipment emergency repair knowledge graph through updated multimodal data, new equipment entities, new fault entities (such as uncovered niche equipment models, new fault types), and new emergency repair resource entities (such as new tools and spare parts) can be extracted from the emergency repair application data, thus improving the entity attributes in the power equipment emergency repair knowledge graph. For the deviations in the association relationships reflected in fault diagnosis errors (such as the incorrect association of "equipment A - fault B"), the entity association relationships in the knowledge graph can be corrected in conjunction with the actual fault handling logic. The entity attribute information in the power equipment emergency repair knowledge graph can be updated based on the equipment operating status and fault handling effects in the emergency repair application data (such as adjusting the fault occurrence probability and supplementing the adaptation scenarios of emergency repair resources).

[0076] By collecting fault diagnosis error cases and feedback optimization suggestions from actual power equipment emergency repair scenarios, the fine-tuned model is supplemented into the sample multimodal data to obtain updated multimodal data. Based on this updated multimodal data, the power equipment emergency repair knowledge graph is optimized, which can enrich the scenario coverage and authenticity of the model training data, make up for the special scenarios and edge cases of actual emergency repairs that were not covered by the original training data, and thus improve the model's adaptability to complex and diverse actual emergency repair scenarios.

[0077] This application provides a fine-tuning system, such as... Figure 3 As shown, Figure 3 The fine-tuning system 300 shown includes a processor 301 and a memory 303. The processor 301 and the memory 303 are connected, for example, via a bus 302. Optionally, the fine-tuning system 300 may also include a transceiver 304. It should be noted that in practical applications, the transceiver 304 is not limited to one type, and the structure of this fine-tuning system 300 does not constitute a limitation on the embodiments of this application.

[0078] Processor 301 may be a CPU (Central Processing Unit), a general-purpose processor, a DSP (Digital Signal Processor), an ASIC (Application Specific Integrated Circuit), an FPGA (Field Programmable Gate Array), or other programmable logic devices, transistor logic devices, hardware components, or any combination thereof. It can implement or execute the various exemplary logic blocks, modules, and circuits described in conjunction with the disclosure of this application. Processor 301 may also be a combination that implements computational functions, such as including one or more microprocessor combinations, a combination of a DSP and a microprocessor, etc.

[0079] Bus 302 may include a pathway for transmitting information between the aforementioned components. Bus 302 may be a PCI (Peripheral Component Interconnect) bus or an EISA (Extended Industry Standard Architecture) bus, etc. Bus 302 can be divided into address bus, data bus, control bus, etc. For ease of representation, Figure 3 The symbol is represented by only one line, but this does not mean that there is only one bus or one type of bus.

[0080] The memory 303 may be a ROM (Read Only Memory) or other type of static storage device capable of storing static information and instructions, RAM (Random Access Memory) or other type of dynamic storage device capable of storing information and instructions, or an EEPROM (Electrically Erasable Programmable Read Only Memory), CD-ROM (Compact Disc Read Only Memory) or other optical disc storage, optical disc storage (including compressed optical discs, laser discs, optical discs, digital universal optical discs, Blu-ray discs, etc.), magnetic disk storage media or other magnetic storage devices, or any other medium capable of carrying or storing desired program code in the form of instructions or data structures and accessible by a computer, but not limited thereto.

[0081] The memory 303 is used to store application code that executes the solution of this application, and its execution is controlled by the processor 301. The processor 301 is used to execute the application code stored in the memory 303 to implement the content shown in the foregoing method embodiments.

[0082] The fine-tuning system includes, but is not limited to: mobile terminals such as mobile phones, laptops, digital radio receivers, PDAs (personal digital assistants), PADs (tablet computers), PMPs (portable multimedia players), and in-vehicle terminals (such as in-vehicle navigation terminals), as well as fixed terminals such as digital TVs and desktop computers. It can also be used for servers, etc. Figure 3 The fine-tuning system shown is merely an example and should not impose any limitations on the functionality and scope of use of the embodiments of this application.

[0083] This application provides a computer-readable storage medium storing a computer program that, when run on a computer, enables the computer to execute the corresponding content in the aforementioned method embodiments.

[0084] This application provides a computer program product including a computer program that, when executed by a processor, implements the methods described in any of the above embodiments.

[0085] It should be understood that although the steps in the flowcharts of the accompanying figures are shown sequentially as indicated by 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 accompanying figures may include multiple sub-steps or multiple stages. These sub-steps or stages are not necessarily completed at the same time, but can be executed at different times, and their execution order is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the sub-steps or stages of other steps.

[0086] The above description is only a partial embodiment of this application. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of this application, and these improvements and modifications should also be considered within the scope of protection of this application.

Claims

1. A method for fine-tuning a power repair model based on a knowledge graph, characterized in that, include: Collect sample multimodal data in the field of power grid operation, maintenance and emergency repair, and classify and label the sample multimodal data to obtain classified and labeled sample multimodal data to be processed. The sample multimodal data includes text data, image data and structured data. The sample multimodal unprocessed data is subjected to categorized data management and control processing to obtain effective multimodal data for management and control, and a knowledge graph for emergency repair of power equipment is constructed based on the effective multimodal data for management and control. Identify the demand indicators corresponding to the effective data of the multimodal management and control, and determine the candidate models and the suitability evaluation system based on the demand indicators; Based on the fitness evaluation system, a standardized test is performed on each candidate model to obtain the comprehensive fitness score corresponding to each candidate model, and the candidate model with the highest comprehensive fitness score is used as the target base model. A hierarchical training dataset is constructed based on the power equipment emergency repair knowledge graph. The hierarchical training dataset includes a general power knowledge dataset, a fault diagnosis special dataset, and an emergency repair practice dataset. Identify the triplet data in the knowledge graph of power equipment emergency repair, and perform fusion processing based on the triplet data and the effective data of the multimodal management and control to form training sample pairs; Based on the training sample pairs and the hierarchical training dataset, the target base model is fine-tuned hierarchically to obtain the fine-tuned model.

2. The power repair model fine-tuning method based on a knowledge graph according to claim 1, characterized in that, The construction of a knowledge graph for emergency repair of power equipment based on the aforementioned effective multimodal data includes: The core entities in the field of power equipment emergency repair are extracted from the effective data of the multimodal management and control. The core entities include equipment entities, fault entities, emergency repair resource entities, and process entities. The relationships between core entities are determined based on the business logic of the power equipment emergency repair field. A knowledge graph storage architecture is built based on a pre-defined graph database, and a triplet structure with entity-attribute-relationship as the core is constructed to form a knowledge system for emergency repair of power equipment. Establish the association and mapping relationship between the power equipment emergency repair knowledge system and the effective multimodal data of management and control to obtain the power equipment emergency repair knowledge graph.

3. The power repair model fine-tuning method based on a knowledge graph according to claim 1, characterized in that, The step of performing hierarchical fine-tuning on the target base model based on the training sample pairs and the hierarchical training dataset to obtain the fine-tuned model includes: The target base model is fine-tuned by embedding domain knowledge in multiple rounds based on the general power knowledge dataset in the hierarchical training dataset, and the accuracy of the validation set after each round of domain knowledge embedding fine-tuning is collected. When the accuracy of the validation set after any round of domain knowledge embedding fine-tuning is higher than the preset validation accuracy threshold, the first basic model is obtained. Based on the fault diagnosis special dataset in the hierarchical training dataset, the first basic model is fine-tuned by focusing on fault features in multiple rounds, and the fault identification accuracy after each round of fault feature focusing fine-tuning is collected. When the fault identification accuracy after any round of fault feature focusing fine-tuning is higher than the preset identification accuracy, a second basic model is obtained. Based on the emergency repair operation dataset in the hierarchical training dataset, the second basic model is fine-tuned through multiple rounds of cross-modal fusion operation, and the recommended accuracy after each round of cross-modal fusion operation fine-tuning is collected. When the recommendation accuracy after any round of cross-modal fusion practical fine-tuning is higher than the preset recommendation accuracy, the fine-tuned model is obtained.

4. The power repair model fine-tuning method based on a knowledge graph according to claim 3, characterized in that, After collecting the fault identification accuracy rate after fine-tuning the focus of each round of fault feature collection, the following is also included: A fault diagnosis-specific dataset is obtained by focusing on and fine-tuning the fault entities based on the fault features of each round. When the entity similarity between fault entities contained in the fault diagnosis special dataset is higher than a preset similarity threshold, the number of times the fault feature is focused and fine-tuned and the fault identification accuracy after each round of fault feature focused and fine-tuned are identified, and the accuracy convergence rate is determined based on the number of times the fault feature is focused and fine-tuned and the fault identification accuracy after each round of fault feature focused and fine-tuned. When the fault identification accuracy after any round of fault feature focusing and fine-tuning is higher than the preset identification accuracy, and the accuracy convergence rate is higher than the preset convergence rate threshold, the second basic model is determined to be generated.

5. The method for fine-tuning a power emergency repair model based on knowledge graphs according to claim 4, characterized in that, When the number of focus adjustments for the fault feature is higher than the preset number of focus adjustments, the following is also included: The fault feature discrimination vector of each highly similar fault entity in the fault diagnosis special dataset is determined based on the knowledge graph of the power field. Each highly similar fault entity in the fault diagnosis special data is combined to obtain multiple highly similar fault entity pairs, and the mean feature discrimination value is determined based on the feature discrimination value of each highly similar fault entity pair. If the mean feature discrimination score is lower than the preset discrimination score threshold, the initial focusing weight of the fault feature focusing fine-tuning corresponding fault feature attention mechanism is adjusted according to the discrimination score difference between the mean feature discrimination score and the preset discrimination score threshold to obtain the target focusing weight; Subsequent fault feature focusing fine-tuning is performed based on target focusing weights.

6. The method for fine-tuning a power emergency repair model based on knowledge graphs according to claim 1, characterized in that, After obtaining the fine-tuned model, the process also includes: Using the fine-tuned model as the teacher model, the corresponding student model is guided by the teacher model to learn features. The intermediate layer feature vectors output by the teacher model and the feature vectors output by the corresponding network layers of the student model are input into a preset distillation loss function to calculate the feature differences and obtain the total loss value of distillation training. The student model is determined based on the demand index. The contribution of each layer of the student model to the emergency repair task of power equipment is analyzed. The importance score of each layer parameter of the student model is calculated based on the summation of the absolute values ​​of the parameter gradients. The network structure of the student model is pruned using a preset iterative pruning strategy. Adaptive quantization is performed on the pruned student model to obtain a quantized optimized model. The quantized optimized model is then deployed and adapted to complete the dynamic compression and full-process optimization of the fine-tuned model.

7. The method for fine-tuning a power emergency repair model based on knowledge graphs according to claim 1, characterized in that, Also includes: Collect emergency repair application data of the fine-tuned model in actual power equipment emergency repair scenarios, including fault diagnosis error cases and feedback optimization suggestions; The emergency repair application data is added to the sample multimodal data to obtain updated multimodal data, and the power equipment emergency repair knowledge graph is optimized based on the updated multimodal data.

8. A fine-tuning system, characterized in that, The fine-tuning system includes: At least one processor; Memory; At least one application, wherein the at least one application is stored in memory and configured to be executed by at least one processor, the at least one application being configured to: execute a knowledge graph-based power emergency repair model fine-tuning method according to any one of claims 1-7.

9. A computer-readable storage medium, characterized in that, include: The computer program contains a knowledge graph-based power emergency repair model fine-tuning method that can be loaded by a processor and executed as described in any one of claims 1-7.

10. A computer program product, characterized in that, The method includes a computer program that, when executed by a processor, implements the steps of the knowledge graph-based power emergency repair model fine-tuning method according to any one of claims 1-7.

Images