Power information questioning system based on dynamic expansion of reference knowledge graph

Abstract

The invention relates to an electric power information questioning system based on reference knowledge graph dynamic expansion. The system comprises a reference knowledge graph construction module, a multi-modal data fusion module, a questioning inertia library generation module and a user questioning analysis module. And deep upgrading is carried out through multi-source data fusion. And the multi-modal data fusion module dynamically weights different modal data through an attention mechanism, and realizes image-text-operation behavior conjoint analysis in combination with cross-modal transfer learning, so that the equipment fault identification efficiency is shortened from manual processing. Simulation fault data of novel equipment is generated through a data enhancement algorithm, a knowledge graph and a question inertia library are automatically updated in combination with incremental learning, a closed-loop evolutionary mechanism of data acquisition-evaluation-enhancement-application is formed, the adaptation period of a new technology is shortened, and the efficiency is improved. And the knowledge graph data quality is continuously optimized through a three-dimensional evaluation system (integrity, accuracy and consistency).

Description

TECHNICAL FIELD

[0001] The technical field relates to power informatization and artificial intelligence technology, in particular to a power information query system based on dynamic expansion of a benchmark knowledge graph. BACKGROUND

[0002] With the rapid development of the field of power information management, the scale of the power grid is continuously expanding, and the demand of users for power information query is increasingly diversified. However, the existing knowledge graph technology has problems such as staticization and insufficient utilization of topological relations in power information processing, and it is difficult to adapt to the dynamic changes of the power grid. Especially in the face of fuzzy queries, the system lacks effective processing mechanisms, which limits the accuracy and response efficiency of the answers. According to statistics, the number of abnormal data queries in the current power information management system reaches the order of magnitude of 100,000 per month, and the query failure rate caused by insufficient utilization of topological relations reaches the order of magnitude of 10,000. The reason for this result is that the existing system's knowledge graph updating mechanism is imperfect and cannot reflect the dynamic changes of the power grid nodes in real time. In addition, when dealing with complex power information queries, grassroots personnel have difficulties in cross-system data flow, repeated queries, semantic understanding deviation, and other problems, and the priority and intent of user queries are not effectively analyzed, resulting in some urgent queries being unable to be processed in a timely manner. Therefore, it is urgent to build a power information query system that integrates benchmark knowledge graph dynamic expansion and power grid node topological relation analysis to realize accurate answering and efficient knowledge processing. For the above problems, although the existing technology has proposed various solutions, it still has limitations in knowledge reasoning and dynamic expansion capabilities in complex scenarios. For example, the patent with the publication number CN112527778A discloses a knowledge graph expansion method based on incremental updating, which improves the dynamic nature of the knowledge graph through multi-source data fusion, but lacks targeted optimization in power grid topological relation analysis; the patent with the publication number CN110991784A discloses a topological relation modeling method based on graph neural networks, which can preliminarily analyze the power grid structure, but still has room for improvement in fuzzy query processing and multi-modal data fusion. Although the above two methods improve the performance of the system to some extent, the ability to analyze fuzzy problems in complex scenarios is still insufficient. With the development of power big data technology and the popularization of smart grids, a large amount of multi-source heterogeneous data can be presented in an electronic way, which provides the possibility for the intelligent upgrading of the power information query system. SUMMARY

[0003] Therefore, the present application aims to provide a load resource data management system based on quality checking rules.

[0004] In order to solve the above technical problems, the technical scheme of the present application is: a power information question system based on benchmark knowledge graph dynamic expansion, comprising a benchmark knowledge graph construction module, a multi-modal data fusion module, a question inertia library generation module and a user question analysis module;

[0005] The benchmark knowledge graph construction module is configured with a federated learning algorithm, a data evaluation algorithm and a data enhancement algorithm, the federated learning algorithm integrates multiple data sources to complete data integration to obtain a benchmark multi-source database, the data evaluation algorithm is used to evaluate the data in the benchmark multi-source database, and the data enhancement algorithm uses an adversarial generative network to enhance the data whose quality is lower than a preset benchmark quality value to update the benchmark multi-source database;

[0006] The multi-modal data fusion module is used to index the data of the benchmark multi-source database with multi-modal information, and the attention values of each data are configured by assigning different modal data weights through an attention mechanism;

[0007] The question inertia library generation module comprises a historical question and answer indexing unit, an inertia generation unit and an inertia library construction unit, the historical question and answer indexing unit is used to index the historical question and answer situation to generate an indexing feature vector, the inertia generation unit is used to cluster analyze the clue clusters of the indexing feature vector to generate inertia mean vectors corresponding to different clue sub-clusters, and the inertia library construction unit obtains an expansion sub-word library of each inertia mean vector from the benchmark multi-source database, the expansion sub-word library comprises several expansion keywords;

[0008] The user question analysis module comprises an inertia expansion unit, an entity expansion unit and an expansion evaluation unit, the inertia expansion unit obtains user information of the current question to generate an evaluation vector, matches the corresponding inertia mean vector according to the user evaluation vector to configure the corresponding expansion keyword, and supplements the expansion keyword into the current question through a semantic adaptability strategy, the entity expansion unit is configured with an entity filling strategy, the entity filling strategy is used for the entity relationship of the current question, and the entity key word is supplemented into the current question according to the entity relationship through the user evaluation vector, and the expansion evaluation unit is configured with an expansion evaluation strategy for calculating an expansion reliable vector of the current question to replace the user evaluation vector, and the current question is sent to the inertia expansion unit or the entity expansion unit again until the expansion reliable vector meets a preset reliability condition.

[0009] Further, the baseline knowledge graph construction module comprises an entity relationship extraction unit and a topological relationship modeling unit; the entity relationship extraction unit adopts an improved BERT-GNN fusion model for entity and relationship extraction, first extracts named entities in the text using the BERT model, then inputs the entity information into the GNN to mine the hidden relationship between entities, and performs clustering analysis according to the generated entity relationship to obtain entity relationship triples with different relationship characteristics; the topological relationship modeling unit constructs a dynamic power grid node topological relationship model based on a spatio-temporal graph convolution network, updates the weights of node attributes and edges in real time through spatio-temporal convolution operation, and simultaneously creates a virtual topological model in combination with a digital twin algorithm.

[0010] Further, the multi-modal data fusion module comprises a modal screening unit and a fusion calculation unit; the modal screening unit is used for indexing the data of the baseline multi-source database according to the relationship characteristics, and deleting data without relationship characteristic indexing; the fusion calculation unit assigns weights to different modal data based on the information entropy value of the modal data by using an attention mechanism, and the lower the information entropy, the higher the weight of the modal data, wherein the information entropy is obtained through a preset entropy generation algorithm.

[0011] Further, the history question and answer indexing unit is configured with a topological calculation sub-algorithm, a distance calculation sub-algorithm, a time calculation sub-algorithm, and a category calculation sub-algorithm, the topological calculation sub-algorithm is used for calculating a topological relationship component, the distance calculation sub-algorithm is used for generating a distance relationship component, the time calculation sub-algorithm is used for generating a time calculation component, and the category calculation sub-algorithm is used for generating a category calculation component, and the indexing feature vector is generated through the topological relationship component, the distance relationship component, the time calculation component, and the category calculation component.

[0012] Further, the inertia generation unit is configured with a plurality of clue clusters, different indexing feature vectors are assigned to different clue clusters according to user question and answer text information, and clustering analysis is performed on each clue cluster through the indexing feature vector to obtain different clue sub-clusters and calculate the inertia mean vector of each clue sub-cluster according to the indexing feature vector.

[0013] Further, the inertia library construction unit extracts user question and answer text information through a preset text information extraction strategy to obtain the basic keywords of each clue sub-cluster, and obtains associated keywords from the baseline multi-source database to construct an expanded sub-word library of the clue sub-cluster by taking the basic keywords as indexes.

[0014] Further, the inertia expansion unit obtains a vector missing deviation by subtracting the user evaluation vector from the inertia average vector, and obtains all qualified expansion keywords from the expansion sub-keyword library indexed by the vector missing deviation, the semantic adaptability strategy is configured with a semantic evaluation constraint, and the expansion keywords meeting the semantic evaluation constraint are supplemented into the current question.

[0015] Further, the entity filling strategy is configured with an entity verification relationship, and the current question is verified through the entity verification relationship to obtain a missing entity class, and the qualified entity verification relationship is called from the benchmark multi-source database through the missing entity class.

[0016] Further, the expansion evaluation strategy is also configured with a rejection criterion, and when the expansion evaluation vector of the current question meets the rejection criterion, the current question is rejected.

[0017] Further, the benchmark knowledge graph construction module establishes a data quality evaluation system when collecting multi-source data, and evaluates from three dimensions of data integrity, accuracy and consistency; the data integrity is judged by checking the field missing condition, the accuracy is evaluated by cross verification and comparison with the authoritative data source, and the consistency is judged by checking whether the same data in different data sources is the same; the benchmark knowledge graph construction module performs targeted processing on the data according to the evaluation results, supplements the missing field if the integrity is insufficient, corrects the error data if the accuracy is not up to standard, and unifies the data expression if there is a problem with the consistency.

[0018] The technical effects of the present application mainly embody in the following aspects: dynamic knowledge modeling capability breakthrough, through the space-time graph convolution network and digital twin algorithm of the benchmark knowledge graph construction module, realizing minute-level topological relationship dynamic update, integrating multi-source data through federated learning, and completely solving the hysteresis problem of static atlas. Fuzzy question processing efficiency is improved, through the cooperation of the question inertia library generation module and the user question analysis module, the keywords are automatically expanded based on the historical question mode, the missing entities are supplemented by the entity filling strategy, the fuzzy question analysis accuracy is improved to more than 85%, and the interaction times are reduced. Multi-source data fusion depth is upgraded. The multi-modal data fusion module dynamically weights different modal data through the attention mechanism, realizes the joint analysis of "image-text-operation behavior" through cross-modal transfer learning, and shortens the device fault identification efficiency from manual processing. Through the data enhancement algorithm, the simulation fault data of new devices is generated, the knowledge graph and the question inertia library are automatically updated through incremental learning, forming a closed-loop evolution mechanism of "data collection-evaluation-enhancement-application", the adaptation cycle of new technologies is shortened, and the data quality of the knowledge graph is continuously optimized through the three-dimensional evaluation system (integrity, accuracy, consistency). BRIEF DESCRIPTION OF DRAWINGS

[0019] Figure 1It is the overall module structure schematic diagram of the system of the application.

[0020] Figure 2 It is the detailed structure diagram of the reference knowledge graph construction module.

[0021] Figure 3 It is the working principle diagram of the multi-modal data fusion module.

[0022] Figure 4 It is the function block diagram of the question inertia library generation module.

[0023] Figure 5 It is the running logic diagram of the user question analysis module. DETAILED DESCRIPTION

[0024] The specific embodiments of the application are further described in detail below in conjunction with the accompanying drawings, so that the technical scheme of the application is easier to understand and master.

[0025] The application relates to a power information question system based on dynamic expansion of a reference knowledge graph, and the overall module structure is shown in the figure. Figure 1 The system comprises a reference knowledge graph construction module, a multi-modal data fusion module, a question inertia library generation module, a user question analysis module and a monitoring and maintenance module. These modules are connected through data flow to form a complete closed-loop system, wherein the reference knowledge graph construction module provides basic data support for the subsequent modules, the multi-modal data fusion module performs screening and fusion processing on the input data, the question inertia library generation module generates a question template based on the fused data, the user question analysis module combines the question inertia library and the knowledge graph to complete question analysis, and the monitoring and maintenance module is responsible for real-time monitoring of the running state of each module and maintenance.

[0026] The benchmark knowledge graph construction module is configured with a federated learning algorithm, a data evaluation algorithm, and a data enhancement algorithm. The federated learning algorithm integrates data from multiple data sources to obtain a benchmark multi-source database. The data evaluation algorithm is used to evaluate the data in the benchmark multi-source database. The data enhancement algorithm uses an adversarial generative network to enhance the data whose quality is lower than the preset benchmark quality value to update the benchmark multi-source database. The key algorithms are described in detail as follows: 1. Federated learning algorithm: The application mode is to use a horizontal federated learning framework to integrate data from power grid companies, equipment manufacturers, and operation and maintenance units without sharing original data. For example, power companies in different regions can jointly train knowledge graphs on a federated server while keeping their local data private. The data processing flow is as follows: data preprocessing (each participant independently cleanses data), feature alignment (unifies entity coding rules), distributed training (each node trains a local model), model aggregation (federal server fuses model parameters), and benchmark database generation (integrates multi-party knowledge). 2. The data evaluation algorithm first establishes a three-dimensional evaluation system. The first aspect is completeness evaluation, which is achieved by field missing rate statistics (such as marking as low completeness if the missing rate of device parameters exceeds 30%) and rule-based reasoning to complete missing fields (such as automatically filling in the rated voltage according to the device type). The second aspect is accuracy evaluation, which is achieved by using the cross-validation method to compare the data with the authoritative data source (such as the State Grid Equipment Standard Library), and triggering the manual review process if the error of a field exceeds 5%. The third aspect is consistency evaluation, which is achieved by using a semantic hashing algorithm to detect differences in the representation of the same entity in different data sources (such as "transformer" and "power transformer"), and normalizing based on the domain ontology. The benchmark quality value is generated by weighting the completeness, accuracy, and consistency. 3. The data enhancement algorithm is a generative adversarial network (GAN) application: a GAN model specific to the power field is constructed to address the problem of small sample data missing. For example, to address the lack of fault data for new intelligent devices, the generator learns the distribution characteristics of existing fault data to generate simulated fault samples (such as a virtual case of "smart circuit breaker communication anomaly"), which are verified by the discriminator and supplemented to the benchmark database. The enhancement strategy uses conditional GAN (cGAN) to generate corresponding data samples based on entity types, ensuring the relevance of the enhanced data to the domain. For example, inputting the "transformer" category, the generator can output various fault feature data of the device. It should be noted that the combination of the three algorithms improves the data format uniformity while ensuring data integrity and filtering out unnecessary data.

[0027] The specific structure of the benchmark knowledge graph construction module is shown in Figure 2 which includes an entity relationship extraction unit and a topological relationship modeling unit.

[0028] The entity relationship extraction unit uses an improved BERT-GNN fusion model to process multi-source heterogeneous data in the power industry. In actual operation, first, the text data from different data sources is input into the BERT model to extract named entities and their context features, and then these features are passed to the GNN model to mine the hidden relationships between entities. For example, when processing a document about substation equipment, the BERT model can identify key entities such as "transformer" and "circuit breaker", and the GNN model further analyzes the electrical connection relationships between these entities. The technical architecture adopts a three-level processing architecture of "BERT model + GNN network + clustering analysis" to automatically extract entities and relationships for multi-source heterogeneous data in the power industry, such as equipment manuals, maintenance records, and dispatch logs. The core algorithm is as follows: 1. BERT entity extraction layer: based on a pre-trained model in the power industry (integrating professional term vectors such as transformer and circuit breaker), the bidirectional Transformer is used to capture named entities in the text. For example, when processing the text "220kV substation main transformer oil temperature abnormality", entities such as "220kV substation", "main transformer", and "oil temperature" can be accurately identified, and entity feature vectors containing position information are generated. 2. GNN relationship mining layer: input the entity features output by BERT into the graph neural network, with entities as nodes and context semantics as edges, and use graph convolution operations (such as GCN or GAT) to mine hidden relationships. For example, from "circuit breaker tripping causes line outage", the causal relationship of "circuit breaker - tripping - line" is extracted, and the confidence of different relationships is weighted through the attention mechanism. Triple clustering generation: using density clustering algorithms such as DBSCAN, according to the semantic similarity and topological correlation of entity relationships, the same type of relationships is aggregated into entity relationship triples (subject - relationship - object). For example, "transformer overload" and "line overload" are aggregated into the abstract triple category of "equipment - operating status - abnormality". When processing substation fault reports, this unit can extract "circuit breaker (entity) - fault type (relationship) - refusal to act (object)" from unstructured text, and generate "refusal to act - possible cause - mechanism jamming" extended relationships based on historical data to form a complete fault knowledge chain.

[0029] The topology relationship modeling unit constructs a dynamic power grid node topology relationship model based on a spatio-temporal graph convolution network. In this process, the system collects real-time operation data of the grid nodes, updates the node attributes and edge weights using spatio-temporal convolution operations, and creates a virtual topology model using a digital twin algorithm. For example, when a power transmission line fails, the system can adjust the attribute values of the relevant nodes and edge weights in real time, thereby reflecting the changes in the actual power grid. In addition, the entity relationship extraction unit is also configured with a knowledge distillation algorithm, which is used to extract key knowledge from large amounts of unstructured data and map it to structured data. For example, by analyzing historical maintenance records, the system can extract common causes of equipment failure and corresponding solutions and store them in the knowledge graph.

[0030] The specific technical framework is a hybrid modeling system based on "spatio-temporal graph convolution + digital twin + dynamic update", which constructs the topology relationship network of the grid nodes in real time. The spatio-temporal graph convolution network instance: taking the physical nodes of the power grid (such as substations and transmission lines) as graph nodes and the electrical connection relationship as edges, the spatio-temporal convolution kernel (such as ST-GCN) is used to process the spatial topology features (node connection strength) and time series features (load fluctuation, failure frequency) simultaneously. For example, when the load rate of a line exceeds the threshold, the system calculates the influence weight of the node on the adjacent nodes through spatio-temporal convolution and updates the edge weight of the topology graph. The digital twin fusion method is: combining power grid three-dimensional modeling and real-time monitoring data to create a virtual topology model. Through the twin model, the running state of the power grid is simulated, for example, in the maintenance scenario, the virtual model can simulate the load transfer of other lines after a certain line is disconnected, providing a prediction basis for topology relationship update. The dynamic update mechanism sets the topology update trigger conditions (such as equipment state change, dispatching instruction issuance), and adjusts the node attributes (such as voltage level, load state) and edge weights (such as transmission capacity, failure probability) in real time through incremental learning algorithms (such as OnlineGCN). For example, when the switch state of a substation changes, the system completes the recalculation of the topology relationship in a shorter time. For example: in power grid dispatching, this unit can construct a topology network containing "substation-line-load center", display the power flow distribution of each node in real time, and assist in fault location through a topology distance algorithm (such as shortest path priority), thereby shortening the traditional fault troubleshooting time. The knowledge distillation algorithm is as follows: unstructured knowledge extraction: for a large amount of text data in the power field (such as maintenance reports, academic papers), the distillation technology is used to map the unstructured knowledge into structured triples. For example, from the report "the main transformer of a substation tripped due to winding overheating", the knowledge unit "main transformer-failure cause-winding overheating" is extracted. The distillation process: text segmentation, semantic role labeling, key information extraction, knowledge mapping (mapped to the knowledge graph nodes through the domain ontology), confidence evaluation (knowledge below the threshold needs to be manually checked and re-labeled or supplemented).

[0031] The function of the benchmark knowledge graph construction module is to construct a benchmark database containing core knowledge such as power equipment, topological relationship, and operation rules, to improve data coverage and entity relationship accuracy, and to provide a unified semantic standard for multi-modal data fusion. For example, when the multi-modal module processes the voice question "transformer temperature anomaly", it can directly associate the "transformer-temperature-threshold" relationship in the knowledge graph to quickly locate the fault knowledge. Through a dynamic updating mechanism (minute-level topology refresh, hour-level data enhancement), the knowledge graph is ensured to reflect the real-time operation state of the power grid, solving the problem that traditional static graphs cannot adapt to the dynamic changes of the power grid. For example, when new equipment is connected to the power grid, the module automatically updates the equipment topological relationship and synchronizes it to the question inertia library, enabling the system to respond to related queries in real time. The module provides a knowledge base for entity expansion and relationship reasoning for the user question analysis module. For example, when the user asks "the reason for the outage of a certain line", the entity filling strategy can automatically supplement the missing entity keywords based on the "line-circuit breaker-trip" relationship in the knowledge graph, converting the fuzzy question into a precise query "whether the circuit breaker of a certain line tripped due to overload". Through knowledge distillation and data enhancement, the implicit knowledge in the graph (such as equipment fault chain reaction and dispatching operation specification) is enriched, and the reasoning ability of the system for complex problems is improved. For example, when processing the question "how to avoid transformer station busbar failure", the system can generate a composite answer containing fault prevention measures based on the "busbar-protection configuration-backup automatic switching" relationship in the graph. To address the complexity of the power grid topology, a high-precision power grid topology model is constructed using spatio-temporal graph convolution and digital twin technology, supporting question optimization based on topological relationships (such as "the status of the line within 10 kilometers of a certain transformer station"). Compared with traditional retrieval methods, the efficiency is improved. Using federated learning and data evaluation system, multi-source heterogeneous data (such as SCADA real-time data, PMS equipment account, and weather data) is integrated to solve the problem of scattered data and non-uniform standards in the power system, providing a unified data interface for cross-system queries. For example, users can obtain the associated information of "certain regional weather-line load-equipment status" through a single question without manually searching across systems. The benchmark database output by the module serves as the data source for the generation of the question inertia library, and through historical question and answer indexing and clustering analysis, the expansion strategy for user questions is continuously optimized. For example, when a certain type of equipment failure query frequently occurs, the module automatically enhances the related knowledge of that equipment and updates the expansion keywords in the inertia library, improving the efficiency of subsequent similar question analysis. Through data quality evaluation and enhancement mechanisms, a closed loop of "data collection-evaluation-enhancement-application-feedback" is formed, continuously improving the completeness and accuracy of the knowledge graph, enabling the system to evolve in operation and adapt to the knowledge update needs of the power industry (such as new energy equipment access and smart grid new technology application).

[0032] The multi-modal data fusion module is used for multi-modal information indexing of data of the benchmark multi-source database, and attention mechanism is used to allocate weights of different modal data to configure attention values of each data. Figure 3 As shown in the figure, it contains a modal screening unit and a fusion calculation unit inside. The trigger condition of the attention weight dynamic adjustment algorithm is when the information entropy of a certain modal data changes by more than 20% (such as sudden failure leading to a sharp increase in sensor data entropy value), the weight redistribution mechanism is started; the adjustment strategy is to use the reinforcement learning (RL) framework, taking the knowledge graph query accuracy rate as the reward function, and optimizing the weights of each mode through the Q-Learning algorithm. For example, in the fault handling scene, the system learns from historical data that when the weight of voice instruction is increased to 0.5, the fault positioning accuracy rate is increased by 15%, and then the voice weight in this scene is dynamically adjusted. The cross-modal semantic alignment algorithm is: the feature projection rule is as follows: map different modal features to a unified knowledge graph semantic space (such as using the TransE model to project image features and text features into entity vectors); the consistency loss function first defines the cross-modal loss as:

[0033]

[0034] Where L is the cross-modal feature alignment loss value, used to optimize the semantic consistency of images and texts, v i is the i-th image feature, t i is the i-th corresponding text feature, f(.) is the projection function, and through contrastive learning, the cross-modal feature distance of the same entity is minimized. Euclidean distance square, which measures the difference between feature vectors, the modal missing compensation algorithm is configured as: the prediction model is as follows: when a certain modal data is missing (such as no inspection image uploaded), the prediction data is generated according to other modal features; an example of the compensation strategy is that when infrared thermal imaging data is missing, based on historical temperature data and equipment load text information, the VAE generates a predicted temperature distribution image of the equipment, and marks the predicted area with a gray mask to prompt manual verification.

[0035] The modal screening unit judges the effectiveness of the voice tone data and the operation behavior data. For example, when the user asks a question by voice, the system analyzes the emotional information in the voice and compares it with the emotional expression in the text information. If the two are inconsistent, the voice data is marked as invalid data. Similarly, for the operation interface click behavior data, the system analyzes whether the click frequency and sequence conform to the normal logic. For example, if the user frequently clicks a certain button in a short period of time and the sequence does not conform to the regular operation process, it is considered invalid data. Its technical architecture adopts a double-layer filtering architecture of "relation feature indexing-invalid data elimination", which screens the effectiveness of multi-source modal data (text, voice, image, operation behavior data) in the power field. The relation feature indexing algorithm is implemented in the following way: based on the ontology of the power field, a modal data correlation matrix is constructed, and the semantic similarity calculation (such as Word2Vec+cosine distance) is used to identify the entity relation features contained in the data. For example, for substation inspection image data, the target detection algorithm is used to identify the device type (transformer, circuit breaker), and the "device-location-state" relationship in the knowledge graph is indexed and matched. The invalid data determination rules are as follows: text data: determine invalid by keyword density detection (such as power professional vocabulary proportion <10%) and syntax completeness analysis (more than 2 missing subjects, verbs and objects); voice data: combined with emotional analysis and text translation consistency check (such as anxious voice tone but calm text content); operation behavior data: through time sequence logic check (such as clicking the "fault query" button more than 20 times in 5 minutes is considered abnormal operation). Power application example: when processing the voice instruction "disconnect 110kV A line switch" of the dispatching center, the modal screening unit first converts the text from the voice recognition, and then indexes the relation feature of "110kV A line-switch-disconnect". If the translated text deviates from the entity, such as "110kV B line", or the voice tone does not match the urgency of the instruction (such as calm tone sending urgent disconnect instruction), the data is determined to be invalid and the re-recording mechanism is triggered.

[0036] The fusion computing unit uses attention mechanism to assign weights to different modal data, where the weight assignment is based on the information entropy value of the modal data. For example, when the information entropy of the text data is low and the information entropy of the voice data is high, the system gives higher weight to the text data. In addition, the fusion computing unit also configures a transfer learning algorithm to migrate the model parameters trained in the general field to the power field and fine-tune them with a small amount of labeled data. For example, by introducing a voice emotion analysis model in the general field, the system can quickly adapt to the specific needs of the power field. The architecture of the fusion computing unit is as follows: technical framework: hybrid fusion model based on "information entropy weighting-attention mechanism-transfer learning", realizing semantic alignment and weight assignment of multi-modal data. The information entropy weight calculation method is as follows: the Shannon entropy formula is used to calculate the uncertainty of each modal data, and the formula is:

[0037]

[0038] where H(X) is uncertainty, p(x i ) is the appearance probability of the i-th feature in modal data, and n is the total number of features in modal data (such as the number of words after text segmentation). For example, the keyword "transformer temperature" in text data has a stable appearance frequency and low information entropy, and is assigned a weight of 0.4; while real-time sensor data fluctuates greatly and has high information entropy, and is assigned a weight of 0.2. The multi-head attention mechanism is as follows: a power field-specific attention network is constructed, and multiple sub-modules such as a topological relationship attention head, a time series attention head, and a device type attention head are set. For example, when processing the "substation fault alarm" scene, the topological attention head focuses on the adjacent node relationship of the fault device, and the time attention head analyzes the time sequence law of the fault occurrence. The specific method of cross-modal transfer learning is to use the domain adaptation technology to fine-tune the general multi-modal model (such as CLIP) on the power data set. For example, the pre-trained model has learned the association between the image "transformer" and the text "Transformer", and through fine-tuning, the cross-modal mapping between "transformer oil temperature image-temperature anomaly text" can be further established.

[0039] A specific example is: in the intelligent inspection scene, the fusion computing unit simultaneously processes visible light images (device appearance), infrared thermal images (temperature distribution), and inspection personnel voice annotations ("A phase sleeve heating"). Through information entropy calculation, the infrared image has a high entropy value (weight 0.3) due to the large fluctuation of temperature data, and the voice annotation has a low entropy value (weight 0.4) due to the clear professional terms, and finally a fusion feature vector containing "device location-temperature anomaly-severity" is generated to assist fault diagnosis.

[0040] In summary, the above modules function to uniformly process multi-source data semantics: invalid data (such as semantically contradictory voice text, abnormal operation behavior data) is eliminated by the modal screening unit, and multi-modal data such as text, voice, and images are relationally feature-indexed based on the ontology of the power field, ensuring that different types of data are semantically aligned under the entity-relation framework of the knowledge graph. For example, the "transformer" entity in the inspection image and the "main transformer" in the voice instruction are uniformly mapped to the "main transformer" standard entity through the knowledge graph. Using information entropy weighting and attention mechanisms, dynamic weights are assigned to different modal data (the lower the information entropy, the higher the weight), highlighting key information. For example, infrared thermal imaging data has a high entropy value and a weight of 0.3 due to large temperature fluctuations, while text annotations with clear professional terms have a low entropy value and a weight of 0.4, ultimately generating a fusion feature vector to assist in fault diagnosis. Through transfer learning, the general multi-modal model is adapted to the power field, establishing cross-modal semantic mapping (such as "transformer oil temperature image-temperature anomaly text" correlation), providing a cross-modal knowledge reasoning basis for the user question analysis module, and improving the processing capability of complex problems (such as "analyze transformer faults in combination with infrared images").

[0041] The functional block diagram of the question inertia library generation module is shown in Figure 4 The historical question and answer indexing unit is used to index historical question and answer situations to generate an indexed feature vector,

[0042] The inertia generation unit is used to perform cluster analysis on the clue clusters of the indexed feature vector to generate inertia mean vectors corresponding to different clue sub-clusters. The inertia generation unit is configured with a number of clue clusters, different indexed feature vectors are assigned to different clue clusters according to user question and answer text information, and cluster analysis is performed in each clue cluster by the indexed feature vector to obtain different clue sub-clusters and calculate the inertia mean vector of each clue sub-cluster according to the indexed feature vector. The specific technical framework is a three-layer inertia modeling system based on "clue cluster division-dynamic clustering-mean calculation", which realizes the cluster analysis of the user question mode. Clue cluster allocation mechanism: a clue cluster feature library (containing 20 pre-set clue clusters such as "transformer fault" and "line outage") is constructed, and each cluster is pre-configured with a set of core keywords; the indexed feature vector is assigned to the corresponding clue cluster using semantic similarity matching (such as BERT-Similarity), and the similarity threshold is set to 0.7. For example, the question "high oil temperature of main transformer" and "transformer overheating" are assigned to the "transformer fault" cluster due to a semantic similarity of 0.85. The dynamic clustering algorithm uses DBSCAN density clustering combined with power grid topology constraints, with the formula being:

[0043] Cluster = {p e N ∈ (p0) | |N ∈(p0) ≥ MinPts

[0044] where Cluster is the generated question pattern clustering cluster, p is the indexing feature vector to be clustered (such as the four-dimensional feature vector of a certain historical question), N ∈ (p0) is the neighborhood with p0 as the center and ∈ as the radius, ∈ is the neighborhood radius (dynamically adjusted based on topological distance), MinPts is the minimum sample size (adaptively set according to the cluster size);

[0045] The correlation degree factor of power grid equipment (such as questions in the same substation having a higher clustering priority) is introduced to optimize the clustering result. Inertial mean vector calculation: for each clue sub-cluster, the weighted average method is used to calculate the inertial mean vector.

[0046]

[0047] where w i is the i-th sample weight (dynamically adjusted based on the question frequency and response accuracy), v i is the indexing feature vector; n is the total number of samples. For example, in the "transformer overload" sub-cluster, the weight of the high-frequency question "reason for main transformer overload" is set to 0.9, and the weight of the low-frequency question is set to 0.1, and a comprehensive inertial vector is generated.

[0048] The inertial library construction unit obtains an expansion sub-word library of each inertial mean vector from a benchmark multi-source database, and the expansion sub-word library includes a plurality of expansion keywords; the inertial library construction unit extracts user question and answer text information by a preset text information extraction strategy to obtain basic keywords of each clue sub-cluster, and obtains associated keywords from the benchmark multi-source database to construct an expansion sub-word library of the clue sub-cluster. The technical implementation process is as follows: based on the closed-loop construction process of "keyword extraction-knowledge graph association-sub-word library verification", a domain-adapted expansion sub-word library is generated. Basic keyword extraction: the TF-IDF+TextRank fusion algorithm is used to extract the basic keywords of the clue sub-cluster, and the formula is:

[0049] Score(w)=α·TF-IDF(w)+(1-α)·TextRank(w)

[0050] where Score(w) is the comprehensive importance score of keyword w, a is the balance coefficient which can be selected as 0.6, and the top 10 keywords are extracted as the base words; IDF(w) is the term frequency-inverse document frequency score (measuring the specificity of keywords in the sub-cluster), and TextRank(w) is the graph ranking score (measuring the centrality of keywords in the semantic network). For example, the base keywords such as "circuit breaker", "protective action" and "short circuit" are extracted from the "line trip" sub-cluster. Knowledge graph association expansion: through the entity relationship network of the benchmark knowledge graph, the 3-order relationship is expanded around the base keywords (such as "circuit breaker"→ "circuit breaker type"→ "spring mechanism"→ "mechanism failure"); the importance of the expanded keywords is calculated by using the PageRank algorithm, and the keywords with a score greater than 0.2 are selected to join the expanded sub-word library.

[0051] Sub-word library verification mechanism: a verification set (containing 500 historical questions) is constructed, the semantic coverage of the expanded keywords is calculated (requiring ≥85%), and the perplexity is used to evaluate the representation ability of the sub-word library for new questions, and the formula is:

[0052]

[0053] where the representation perplexity of the sub-word library W for new questions is lower, indicating stronger adaptability; N is the number of questions in the verification set, and w i The i-th word in the question, P(w i |w i-1 , …, w1) is the word sequence generation probability based on the sub-word library; when the perplexity is greater than 150, the sub-word library updating process is triggered.

[0054] The historical question and answer indexing unit is configured with a topology calculation sub-algorithm, a distance calculation sub-algorithm, a time calculation sub-algorithm, and a category calculation sub-algorithm. The topology calculation sub-algorithm is used to calculate a topology relationship component. The distance calculation sub-algorithm is used to generate a distance relationship component. The time calculation sub-algorithm is used to generate a time calculation component. The category calculation sub-algorithm is used to generate a category calculation component. The indexing feature vector is generated through the topology relationship component, the distance relationship component, the time calculation component, and the category calculation component. A "four-dimensional feature extraction-vector fusion" indexing framework is adopted to perform multi-dimensional feature indexing on historical question and answer data (text questions, response content, and interaction logs) in the power field. I. The topology calculation sub-algorithm is as follows: Based on the power grid topology relationship network of the knowledge graph, the topology distance between the device entities involved in the question is calculated (such as the connection path of the substation-line-load point). The graph traversal algorithm (such as Dijkstra) is used to generate the topology relationship component, which is quantitatively represented as a topology distance matrix. For example, when the question is "a substation outlet switch failure", the topology distance between the switch and the adjacent substation is calculated, and a topology feature vector containing 3 layers of adjacent nodes is generated. II. The distance calculation sub-algorithm is as follows: A hybrid distance measurement model is constructed by combining the Euclidean distance and the power grid topology distance, and the formula is:

[0055] D = a · D euclid +(1-a)·D topo

[0056] Wherein, a is the weight coefficient (default 0.3), D euclid is the Euclidean distance of geographical coordinates, and D topo is the topology path distance; the distance relationship component is generated, such as the question "the line within 5 kilometers from a substation", which outputs a hybrid feature containing geographical distance and topology distance. III. The time calculation sub-algorithm is as follows: The time series decomposition technology (such as STL) is used to analyze the time characteristics of the question, and the seasonal component (such as the question regularity during the daily load peak period) and the trend component (such as the question related to the device maintenance period) are extracted; the time calculation component is generated, for example, the question "transformer overload alarm at 3 o'clock in the morning" is mapped to the time feature vector "night-load low period-abnormal alarm". IV. The category calculation sub-algorithm is specifically set as follows: Based on the power field ontology, a problem category tree (8 categories such as fault query, dispatch operation, and device parameter) is constructed, and a hierarchical classification model (such as HMC) is used for category annotation; the category calculation component is generated, such as the question "110kV line protection setting value" is classified into the "relay protection-setting value query" category, and the category confidence vector is output. The feature vector fusion formula is configured as:

[0057] V index =W t ⊙V t +W d ⊙V d+W tm ⊙V tm +W c ⊙V c

[0058] where V t , V d , V tm , V c are topological, distance, time, category component vectors respectively, W is the corresponding weight matrix (optimized by historical question and answer accuracy), and represents vector dot product.

[0059] Data calculation examples are as follows: Topological relationship: Topological distance matrix generation: Take the question "220kV No. 1 main transformer substation switch failure" as an example:

[0060]

[0061] Extract topological relationship from knowledge graph: No. 1 main transformer substation switch, calculate topological distance using Dijkstra algorithm: The direct connection distance between the main transformer and the switch is 1 (adjacent nodes) The topological distance from the main transformer to the adjacent No. 2 main transformer substation is 3 (through 2 lines) Generate a 3x3 topological distance matrix: Vector representation is topological component: [0, 1, 3, 1, 0, 2, 3, 2, 0] (matrix flattening).

[0062] Distance hybrid distance model example: Question "Line within 5 kilometers from No. 3 main transformer substation": Geographical coordinate Euclidean distance calculation: No. 3 main transformer substation coordinates (116.5°, 39.8°), certain line coordinates (116.7°, 39.9°), Topological path distance calculation: The topological path from the No. 3 main transformer substation to the line in the knowledge graph is 3 edges, D topo= 3 (topological distance unit), hybrid distance calculation (a = 0.3): D = 0.3 x 22.3 + 0.7 x 3 = 6.69 + 2.1 = 8.79. Distance relationship component is [22.3, 3, 8.79] (Euclidean distance, topological distance, hybrid distance). Time: Time series decomposition result: Ask "transformer overload alarm at 2025-06-24 03:00", STL decomposes each component (with a period of 24 hours): Trend component: 3 am belongs to the load valley period (trend value -0.6). Seasonal component: The frequency of overload alarm in this period is 0.2 times per day (seasonal index 0.3). Residual component: Sudden overload event (residual value 0.8). Time feature coding: Hour coding: 3 / 24 = 0.125. Date type: weekday = 0.7, weekend = 0.3 (24th is Tuesday). Time calculation component: [-0.6, 0.3, 0.8, 0.125, 0.7]. Category is hierarchical classification result: Ask "110kV B line protection setting value query", category tree hierarchical classification: first-level category: equipment parameter query (confidence 0.95); second-level category: relay protection parameter (confidence 0.92); third-level category: protection setting value query (confidence 0.88), category confidence vector: V c = [0.95, 0.92, 0.88, 0.15, 0.08, 0.12]

[0063] (The first 3 dimensions are target category confidence, and the last 3 dimensions are non-target category confidence)

[0064] Five, four-dimensional feature vector fusion weight matrix and vector operation: suppose the weight matrix of each component is:

[0065] W t = [0.3, 0.3, 0.4], W d = [0.2, 0.3, 0.5], W tm = [0.1, 0.2, 0.3, 0.2, 0.2],

[0066] W c = [0.4, 0.3, 0.2, 0.1]; Each component vector (simplified to a 3-dimensional example): Topological component: V t = [1, 2, 3], distance component: V d = [8.79, 22.3, 3]; Time component: V tm = [-0.6, 0.3, 0.8]; Category component: V c = [0.95, 0.92, 0.88].

[0067] Fusion calculation formula and result: V index = W t ⊙V t + W d ⊙Vd + tm V tm + c V c

[0068] V index = [0.3 x 1, 0.3 x 2, 0.4 x 3] + [0.2 x 8.79, 0.3 x 22.3, 0.5 x 3] + [0.1 x (-0.6), 0.2 x 0.3, 0.3 x 0.8] + [0.4 x 0.95, 0.3 x 0.92, 0.2 x 0.88]

[0069] [0.1 x (-0.6), 0.2 x 0.3, 0.3 x 0.8] + [0.4 x 0.95, 0.3 x 0.92, 0.2 x 0.88]

[0070] V index = [0.3, 0.6, 1.2] + [1.758, 6.69, 1.5] + [-0.06, 0.06, 0.24]

[0071] + [0.38, 0.276, 0.176]

[0072] V index = [2.378, 7.626, 3.116]. The final indexing eigenvector is obtained.

[0073] In summary, the role of the question inertia library generation module in the scheme: through the historical question and answer indexing unit to extract the four-dimensional eigenvector of topology, distance, time, category (such as the topology distance matrix of the question "the fault of the outgoing line switch of a certain transformer substation", and the load cycle characteristics corresponding to the question time), and use the dynamic clustering algorithm to generate the inertia mean vector of the user's question, and describe the high-frequency question mode (such as the typical eigenvector of the "transformer overload" subcluster). Based on the inertia mean vector and the missing deviation of the user's current question vector, the key words with semantic adaptation are selected from the expanded sub-word library (such as the key words "circuit breaker tripping" and "protection action" when the user asks "line outage"), and the fuzzy question is converted into a structured query, which improves the problem analysis accuracy. Through the knowledge graph association expansion to generate the expanded sub-word library (such as "circuit breaker" → "spring mechanism" → "mechanism failure"), and combine the effectiveness of the historical question and answer verification sub-word library to form a closed loop of "question mode clustering-keyword expansion-knowledge verification", which continuously adapts to the knowledge update in the power field (such as the automatic inclusion of the question mode of the new equipment "solid-state circuit breaker" into the inertia library).

[0074] The running logic of the user question analysis module is as follows Figure 5As shown, the user question analysis module includes an inertia expansion unit, an entity expansion unit, and an expansion evaluation unit. The inertia expansion unit obtains user information of a current question to generate an evaluation vector, matches a corresponding inertia mean vector according to the user evaluation vector to configure a corresponding expansion keyword, and supplements the expansion keyword into the current question through a semantic adaptability strategy

[0075] The entity expansion unit is configured with an entity filling strategy for entity relations of the current question, and supplements entity keywords into the current question through the user evaluation vector according to the entity relations. The expansion evaluation unit is configured with an expansion evaluation strategy for calculating an expansion reliable vector of the current question to replace the user evaluation vector, and re-sending the current question to the inertia expansion unit or the entity expansion unit until the expansion reliable vector meets a preset reliability condition. The entity filling strategy is configured with an entity verification relation, and verifies the current question through the entity verification relation to obtain a missing entity class, and calls an entity verification relation meeting a condition from a benchmark multi-source database through the missing entity class. The expansion evaluation strategy is also configured with a rejection criterion, and when the expansion evaluation vector of the current question meets the rejection criterion, the current question is rejected. The entity expansion unit realizes automatic filling of key entities in the question based on a closed loop process of “entity relation verification-missing identification-knowledge retrieval”. The entity verification relation model: a power field entity relation rule library is constructed, including: a “device-failure” relation: transformer → winding short circuit, a “location-device” relation: substation → high-voltage switch, and a “failure-cause” relation: line trip → lightning overvoltage. A rule engine is used to verify the entity integrity of the question. For example, the question “prop line maintenance” lacks a “maintenance type” entity, triggering the filling mechanism. The missing entity class identification algorithm: based on dependency syntax analysis to determine the missing entity type. For example, in the sentence “query a fault handling scheme”, “a fault” is the “fault type” entity to be filled; through the “fault type-handling scheme” relation chain of the knowledge graph, the specific fault names such as “winding short circuit” and “turn-to-turn fault” are located. The entity keyword screening strategy calculates the entity score in combination with the user evaluation vector and the knowledge graph correlation degree, and the formula is:

[0076] Score(e)=β·Sim(V user ,e)+(1-β)·Rel(e,Q)

[0077] Wherein, e: candidate entity keyword (such as "main transformer"); β: weight coefficient (default 0.6), adjust the influence of user preference; Sim: semantic similarity of user vector and entity (range 0-1); Rel: relationship strength between entity and question (based on knowledge graph PageRank score, range 0-1); Example: filter the "power outage maintenance" and "live maintenance" entities with high association with "line maintenance", and the scores are 0.85 and 0.72 respectively. User question "D line protection setting value", entity expansion unit verifies the "line-protection type-setting value" relationship, identifies the missing "protection type" entity (such as "differential protection" "distance protection"), and supplements "D line differential protection setting value".

[0078] The inertia expansion unit calculates the vector missing deviation by subtracting the user evaluation vector and the inertia mean vector, and obtains all qualified expansion keywords from the expansion sub-keyword library indexed by the vector missing deviation. The semantic adaptability strategy is configured with semantic evaluation constraints, and the expansion keywords that meet the semantic evaluation constraints are supplemented into the current question. First, by using the three-layer processing architecture of "user vector generation-inertia matching-semantic completion", the details of the intelligent semantic expansion core algorithm of user question based on the question inertia library are as follows. The user evaluation vector is generated by combining the BERT pre-training model and the user historical interaction data to generate a multi-dimensional feature vector. For example, when the user asks "110kV line A power outage reason", the keyword vector [110kV, line A, power outage, reason] is extracted, and the user's historical query preference (such as the last 3 queries all involve "protection action") is fused to generate an evaluation vector with 10 dimensions, and each dimension value is a feature importance score (0-1). The inertia mean vector matching method is: the cosine similarity is used to calculate the matching degree of the user evaluation vector and the inertia mean vector in the question inertia library, and the formula is:

[0079]

[0080] V user : user evaluation vector (dimension N), representing the semantic features of the current question, μ: inertia mean vector (dimension N), representing the typical features of the historical similar questions, V user · μ: vector dot product, measure the feature coincidence degree; ‖V user ‖, ‖μ‖: vector L2 norm, used for normalization processing data. Specific example: match to the mean vector [line, power outage, circuit breaker, protection action, overload] of "line power outage-protection action" sub-cluster, the similarity calculation result is 0.82 (threshold 0.6). The vector missing deviation calculation process is to calculate the difference between the user vector and the inertia vector dimension by dimension, and generate a missing deviation vector. For example:

[0081] V diff= [0, 0, 0.7, 0.8, 0.5]

[0082] The dimension with a deviation value > 0.5 (such as "circuit breaker" "protective action") is determined as a keyword dimension that needs to be supplemented. The semantic adaptability strategy establishes double constraint conditions as follows: semantic relevance > 0.7 (word vector cosine distance is calculated based on Word2Vec) syntax integrity score > 80 points (based on dependency syntax analysis model) In this way, multiple filling modes of the updated current question can be obtained from a current question. Filter the keywords that meet the conditions from the expanded sub-word library, such as supplementing "circuit breaker overload" "protective action trip" to the original question to generate "110kV line A whether the protection action tripped due to the overload of the circuit breaker". Power application example: the user asks "equipment abnormality of B substation", the inertia expansion unit matches to the "substation-transformer abnormality" sub-cluster, and supplements keywords such as "main transformer oil temperature" and "cooling system" to form the complete query "whether the main transformer oil temperature of B substation is abnormally high due to the abnormality of the cooling system".

[0083] The expansion evaluation unit is configured with an expansion evaluation strategy for calculating an expansion reliable vector of the current question to replace the user evaluation vector, and re-sending the current question to the inertia expansion unit or the entity expansion unit until the expansion reliable vector meets the preset reliability condition. Based on the three-layer evaluation system of "reliable vector calculation-iterative optimization-quality control", the accuracy of the question expansion is ensured.

[0084] The expansion reliable vector is calculated by constructing a four-dimensional evaluation index vector:

[0085] V reliable = [C s ,C g ,C k ,C h ]

[0086] C s : semantic consistency score (semantic similarity of the question before and after expansion, threshold > 0.8)

[0087] C g : syntax correctness score (syntax analysis model score, threshold > 90 points)

[0088] C k : knowledge coverage score (relationship coverage rate of the expanded keywords in the knowledge graph, threshold > 85%)

[0089] C h : historical matching rate score (matching accuracy with historical successful cases, threshold > 75%)

[0090] Calculation example: C s 0.85 is calculated using BERT-Similarity, and Ck By statistically expanding the number of knowledge graph triplets associated with keywords, we get 0.88.

[0091] Iterative optimization mechanism:

[0092] Trigger iteration when any dimension does not meet the threshold: C s If it is insufficient, return to the inertial expansion unit to adjust the semantic relevance of keywords; C k When insufficient, return the entity expansion unit to supplement the knowledge graph associated entity

[0093] The maximum number of iterations is limited to 3 times. If exceeded, manual review will be initiated.

[0094] Elimination criteria rule: Eliminate questions if one of the following conditions is met: C s <0.5 and C k <0.6 (double lack of semantics and knowledge); ambiguous entities exist and cannot be disambiguated (such as "main transformer" has unclear reference and no historical context); invalid question types (such as non-power field questions, repeated and meaningless questions).

[0095] Example in the power industry: A user asks "a certain device is faulty", which is first expanded to "a certain transformer is faulty". Calculate C k =0.7 (does not cover the specific fault mode), returns to supplement "winding overheating", V reliable =[0.85,92,0.88,0.76], which satisfies all thresholds.

[0096] In summary, the above-mentioned module solves the common fuzzy question problems in the power field, such as "line power failure" and "equipment anomaly", and through inertia expansion and entity filling, it converts unstructured questions into structured queries. For example, "Jia transformer substation anomaly" is expanded to "whether No. 1 main transformer of Jia transformer substation causes alarm due to high oil temperature", so that the query accuracy is improved. By using the historical pattern matching of the question inertia library, the user's potential demand is quickly located. For example, when the user frequently queries "circuit breaker refusal" related problems, the module automatically promotes the expansion priority of such keywords, shortening the response time. As a bridge connecting the basic knowledge graph, multi-modal data and the question inertia library, it realizes cross-module knowledge fusion. For example, the entity relationship rules (such as "transformer-cooling system-failure") are obtained from the knowledge graph, the multi-modal data fusion results (such as infrared image temperature features) are used to assist entity selection, and the expansion strategy is optimized based on the question inertia library (such as high-frequency question mode priority matching), driving the system to form a closed-loop process of "question analysis-knowledge retrieval-answer generation", supporting the full-link intelligentization of power information query. Deeply integrate the knowledge in the power field to realize the automatic completion and standard expression of professional terms. For example, the user's colloquial "transformer is broken" is converted into the standard term "transformer failure", and the professional dimension "whether caused by winding short circuit" is supplemented. Based on the power grid topology relationship and equipment operation rules, the scene adaptability of question analysis is improved. For example, after the topology of the power grid is changed (such as line reconnection), the module automatically updates the "topological distance" calculation rules to ensure the accuracy of the analysis of questions such as "fault impact range". Through the iterative optimization mechanism of the expansion evaluation unit, a closed-loop evolution capability of "analysis-evaluation-optimization" is formed. For example, when a new type of device "solid-state circuit breaker" is connected to the power grid, the module analyzes the newly added question mode and automatically updates the inertia library and entity rules, so that the system's adaptation period to new technologies is shortened from the traditional solution. The construction strategy of the basic knowledge graph is optimized in reverse. For example, if it is found that the lack of data of a certain type of device fault causes the accuracy of the analysis to decrease, the module will trigger the data enhancement process of the knowledge graph, and generate simulated fault cases through GAN algorithm to supplement to the database.

[0097] The monitoring and maintenance module is connected with the benchmark knowledge graph construction module, the multi-modal data fusion module, the question inertia library generation module and the user question analysis module, and is used for monitoring the running state of each module in real time. In actual operation, the system monitors the processing speed, resource occupancy and other indicators of each module in real time by setting performance index monitoring points. For example, when the processing speed of a certain module is detected to decrease significantly, the system automatically triggers the fault diagnosis process to locate the fault cause by analyzing the running log and data interaction record. In addition, the monitoring and maintenance module also backs up the data in the system regularly, and can quickly recover according to the backup data when the system loses data. For example, the system backs up the data in the benchmark knowledge graph construction module to the cloud storage every day to ensure data security. The benchmark knowledge graph construction module establishes a data quality evaluation system when collecting multi-source data, and evaluates from three dimensions of data integrity, accuracy and consistency. For example, data integrity is judged by checking the field missing situation, and if some fields are missing, the corresponding data is supplemented; accuracy is evaluated by cross-validation and comparison with authoritative data sources, and if there is wrong data, it is modified; consistency is checked by checking whether the same data is expressed in different data sources, and if there is difference, the data expression is unified. For example, when the name of a certain power transmission line is not consistent in different data sources, the system will unify and standardize it according to the standard of the authoritative data source. The above embodiments describe in detail the composition relationship, connection relationship and cooperation mode of each module and the internal unit of the system, and show how to provide data support through the benchmark knowledge graph construction module, filter and fusion process the input data through the multi-modal data fusion module, generate the question template through the question inertia library generation module, complete the question analysis through the user question analysis module, and ensure the stable operation of the system through the monitoring and maintenance module. Through the cooperative work of these modules, the system can realize efficient processing and accurate answering of power information questions. In order to better enable relevant personnel in the technical field to fully understand and implement the present application, the specific implementation principle of the present application is further supplemented in combination with a specific application scene. In the actual operation of the power grid dispatching center, the system needs to process diversified power information questions from users. For example, a user asks by voice: "Has the equipment of a certain substation appeared abnormal recently?" accompanied by clicking the "substation state query" button in the operation interface. At this time, the system starts the multi-modal data fusion module, and the modal filtering unit judges the effectiveness of the voice tone data and operation behavior data. The voice tone data analysis result shows that the user's question emotion is neutral, consistent with the emotion expressed in the text information, so the voice data is judged to be effective. At the same time, the operation behavior data analysis shows that the user's click frequency and order are consistent with the regular logic, so the operation behavior data is also judged to be effective.Subsequently, the fusion computing unit assigns weights to different modal data using an attention mechanism, where the information entropy of speech data is lower and the information entropy of text data is higher, so the system gives higher weights to text data, and through the transfer learning algorithm, the parameters of the general domain voice emotion analysis model are transferred to the power domain to adapt to specific needs. After completing the multi-modal data fusion, the system enters the question inertia library generation module. The clustering analysis unit first performs noise reduction processing on the extracted question sentence feature sequence, and uses a sliding window filtering algorithm to traverse the feature sequence, dynamically adjusting the window length to ensure the noise reduction effect. Subsequently, the system uses principal component analysis algorithm to reduce the dimension of the noise-reduced feature sequence, and determines the dimension reduction dimension by accumulating the variance contribution rate, and compresses the high-dimensional feature sequence into a low-dimensional vector. In the clustering process, the system introduces the topological distance between the nodes of the power grid as the distance measurement standard, and combines the dynamic time warping algorithm to measure the similarity of the time series features of the question sentences. By performing word segmentation, part-of-speech tagging, and semantic role labeling on the question sentences, the system extracts the order of word appearance and semantic change trend, and converts it into a feature sequence. The generative learning unit generates diversified question templates based on the question patterns extracted by the meta-learner, for example, the generator derives multiple expression forms by replacing synonyms and adjusting the word order, while the discriminator distinguishes the generated templates from the real question data and feeds back the results to the generator. Next, the system enters the user question analysis module. The question understanding unit constructs a multi-turn dialogue understanding model based on Transformer, and combines the power grid domain knowledge graph for question analysis. The system uses the multi-head attention mechanism to capture long-distance dependencies in the question, and uses the interactive attention mechanism to dynamically adjust the understanding focus of the current question according to the user's historical questions. The knowledge matching unit designs a knowledge matching strategy based on hybrid retrieval, combining rule-based retrieval and semantic-based retrieval. The rule-based retrieval performs accurate matching in the knowledge graph based on the keyword "substation" and the syntax structure, while the semantic-based retrieval uses knowledge graph embedding vectors for semantic similarity retrieval. The fuzzy question processing unit constructs an optimized agent, taking the fuzzy question "Has any abnormality occurred in the equipment of the nearest substation?", the question inertia library information, and the knowledge graph knowledge as state input, and generating supplementary information to improve the query sentence. For example, the system generates supplementary information "query time range is the past 7 days" based on the historical question patterns in the question inertia library, and generates reasonable supplementary content "Please provide the substation name or number" based on the generative pre-training model. The optimization agent sets an exploration-exploitation balance mechanism, with a higher probability of exploration operation at the initial stage of system operation, and gradually reduces the exploration probability and preferentially selects the supplementary method with higher historical reward. Throughout the process, the monitoring and maintenance module monitors the running state of each module in real time. The system monitors the processing speed, resource occupancy rate and other indicators of each module in real time by setting performance indicator monitoring points.For example, when the processing speed of a certain module is detected to be significantly reduced, the system automatically triggers the fault diagnosis process to locate the cause of the fault by analyzing the operation log and data interaction record. In addition, the monitoring and maintenance module backs up the data in the benchmark knowledge graph construction module to the cloud storage on a daily basis to ensure data security. The above steps demonstrate the operation principle of the system in actual scenarios. Through the provision of basic data support by the benchmark knowledge graph construction module, the screening and fusion processing of input data by the multi-modal data fusion module, the generation of question templates by the question inertia library generation module, the completion of question analysis by the user question analysis module, and the guarantee of stable operation of the system by the monitoring and maintenance module, efficient processing and accurate answering of power information questions are achieved. The above description is only a preferred embodiment of the present application and is not intended to limit the present application. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principles of the present application shall be included in the protection scope of the present application.

[0098] Of course, the above is only a typical example of the present application, and in addition to this, the present application can have other various specific embodiments, and any technical solutions formed by equivalent replacement or equivalent transformation shall fall within the scope of protection of the present application.

Claims

1. A power information questioning system based on dynamic expansion of a benchmark knowledge graph, characterized by: It includes a benchmark knowledge graph construction module, a multimodal data fusion module, a question inertia library generation module, and a user question parsing module; The benchmark knowledge graph construction module is configured with a federated learning algorithm, a data evaluation algorithm, and a data enhancement algorithm. The federated learning algorithm combines multiple data sources to complete data integration to obtain a benchmark multi-source database. The data evaluation algorithm is used to evaluate the data in the benchmark multi-source database. The data enhancement algorithm uses a generative adversarial network to enhance data whose quality is lower than a preset benchmark quality value to update the benchmark multi-source database. The multimodal data fusion module is used to perform multimodal information indexing on the data of the benchmark multi-source database, and to assign weights to the data of different modalities through the attention mechanism to configure the attention value of each data; The question inertia library generation module includes a historical question and answer indexing unit, an inertia generation unit, and an inertia library construction unit. The historical question and answer indexing unit is used to index historical question and answer situations to generate indexed feature vectors. The inertia generation unit is used to perform cluster analysis on clue clusters of the indexed feature vectors to generate inertia mean vectors corresponding to different clue subclusters. The inertia library construction unit obtains an expanded sub-lexicon for each inertia mean vector from a benchmark multi-source database. The expanded sub-lexicon includes a plurality of expanded keywords. The user question parsing module includes an inertia expansion unit, an entity expansion unit and an expansion evaluation unit. The inertia expansion unit obtains user information of the current question to generate an evaluation vector, matches the corresponding inertia mean vector according to the user evaluation vector to configure the corresponding expansion keywords, and supplements the expansion keywords to the current question through the semantic adaptability strategy. The entity expansion unit is configured with an entity filling strategy, which is used for the entity relationship of the current question, and filters the entity keywords to the current question through the user evaluation vector according to the entity relationship. The expansion evaluation unit is configured with an expansion evaluation strategy for calculating the expansion reliability vector of the current question to replace the user evaluation vector, and resends the current question to the inertia expansion unit or the entity expansion unit until the expansion reliability vector meets the preset reliability condition.

2. The power information questioning system based on dynamic expansion of a benchmark knowledge graph according to claim 1, characterized in that: The benchmark knowledge graph construction module includes an entity relationship extraction unit and a topology relationship modeling unit; The entity relationship extraction unit uses an improved BERT-GNN fusion model to extract entities and relationships. First, the BERT model is used to extract named entities in the text. Then, the entity information is input into the GNN to mine the hidden relationships between entities. Cluster analysis is performed based on the generated entity relationships to obtain entity relationship triplets with different relationship characteristics. The topology relationship modeling unit constructs a dynamic power grid node topology relationship model based on the spatiotemporal graph convolutional network, updates node attributes and edge weights in real time through spatiotemporal convolution operations, and creates a virtual topology model in combination with the digital twin algorithm.

3. The power information questioning system based on dynamic expansion of a benchmark knowledge graph according to claim 2, characterized in that: The multimodal data fusion module includes a modality screening unit and a fusion calculation unit; the modality screening unit is used to index the data of the benchmark multi-source database according to the relationship features and delete the data that does not have the relationship feature indexing; the fusion calculation unit adopts an attention mechanism to assign weights to different modal data based on the information entropy value of the modal data, and the modal data with lower information entropy has higher weight, and the information entropy is obtained by a preset entropy generation algorithm.

4. The power information questioning system based on dynamic expansion of a benchmark knowledge graph according to claim 1, characterized in that: The historical question and answer indexing unit is configured with a topology calculation sub-algorithm, a distance calculation sub-algorithm, a time calculation sub-algorithm and a category calculation sub-algorithm. The topology calculation sub-algorithm is used to calculate and generate a topological relationship component, the distance calculation sub-algorithm is used to generate a distance relationship component, the time calculation sub-algorithm is used to generate a time calculation component, and the category calculation sub-algorithm is used to generate a category calculation component. The indexing feature vector is generated by the topology relationship component, the distance relationship component, the time calculation component and the category calculation component.

5. The power information questioning system based on dynamic expansion of a benchmark knowledge graph according to claim 4, characterized in that: The inertia generation unit is configured with several clue clusters, and different indexing feature vectors are assigned to different clue clusters according to the user question and answer text information. Cluster analysis is performed in each clue cluster through the indexing feature vectors to obtain different clue sub-clusters and the inertia mean vector of each clue sub-cluster is calculated based on the indexing feature vectors.

6. The power information questioning system based on dynamic expansion of a benchmark knowledge graph according to claim 4, characterized in that: The inertia library construction unit extracts user question and answer text information through a preset text information extraction strategy to obtain basic keywords for each clue sub-cluster, and uses the basic keywords as indexes to obtain related keywords from a benchmark multi-source database to construct an expanded sub-lexicon of the clue sub-cluster.

7. The power information questioning system based on dynamic expansion of a benchmark knowledge graph according to claim 1, characterized in that: The inertia expansion unit obtains the vector missing deviation by taking the difference between the user evaluation vector and the inertia mean vector, and obtains all qualified expansion keywords from the expansion sub-lexicon using the vector missing deviation as an index. The semantic adaptability strategy is configured with semantic evaluation constraints, and the expansion keywords that meet the semantic evaluation constraints are added to the current question.

8. The power information questioning system based on dynamic expansion of a reference knowledge graph according to claim 7, characterized in that: The entity filling strategy is configured with an entity verification relationship, and verifies the current problem through the entity verification relationship to obtain a missing entity class, and retrieves an entity verification relationship that meets the conditions from a benchmark multi-source database through the missing entity class.

9. The power information questioning system based on dynamic expansion of a benchmark knowledge graph according to claim 1, characterized in that: The expanded evaluation strategy is further configured with a rejection benchmark. When the expanded evaluation vector of the current problem meets the rejection benchmark, the current problem is rejected.

10. The power information questioning system based on dynamic expansion of a reference knowledge graph according to claim 1, characterized in that: The benchmark knowledge graph construction module establishes a data quality assessment system when collecting multi-source data, and evaluates data from three dimensions: data integrity, accuracy, and consistency; data integrity is judged by checking whether fields are missing, accuracy is assessed by cross-validation and comparison with authoritative data sources, and consistency is assessed by checking whether the representation of the same data in different data sources is the same; the benchmark knowledge graph construction module performs targeted processing on the data based on the assessment results, supplements missing fields if the integrity is insufficient, corrects erroneous data if the accuracy does not meet the standards, and unifies data representation if there are consistency issues.

Images