Power item energy question-answering system and method based on multi-data fusion and semantic analysis

Abstract

The invention belongs to the technical field of electric power systems, and particularly relates to an electric power item energy question answering system and method based on multi-data fusion and semantic analysis. Comprising a multi-source data access module, a data cleaning and standardization module, a data fusion and association module, a semantic analysis and model training module, a multi-round dialogue interaction module and a real-time feedback optimization module which work cooperatively through data circulation and instruction interaction. Text, image and sensor data are collected through a multi-source data access module and are classified and temporarily stored, after the data are processed through a data cleaning and standardization module, a model of three-level nodes and four types of association edges is constructed through a data fusion and association module based on a graph convolutional neural network (GCN), deep association and unified representation of multi-source data are achieved, and the data are classified and temporarily stored. The objective of the invention is to solve the problems of insufficient multi-source data processing, inaccurate semantic understanding, dialogue stiffness and model optimization deficiency of an existing power project intelligent question-answering system.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of power systems, and in particular to a power project Q&A system and method based on multi-data fusion and semantic analysis. BACKGROUND

[0002] With the acceleration of digital transformation of the power industry, power project management has entered the stage of full-process data, and multi-source heterogeneous data such as text (design scheme, operation and maintenance records, etc.), images (equipment inspection photos, construction videos, etc.), and sensor data (equipment operation parameters, environmental monitoring data, etc.) are generated throughout the project life cycle. At the same time, AI technologies such as large language models (LLM), knowledge graphs, and graph neural networks are penetrating into the power field, driving the transformation of project management towards "data-driven + intelligent decision-making", and intelligent Q&A systems, as the core carrier of human-computer interaction, have become a key tool for improving knowledge acquisition efficiency.

[0003] Currently, intelligent Q&A systems in the power field are mainly divided into three categories: first, systems based on rule libraries, which rely on manually preset rules to match questions and output fixed answers; second, systems based on basic large language models, which are trained on general corpus and can only handle simple common sense questions; third, hybrid systems combining rule libraries and basic LLMs (such as some power enterprise project knowledge query systems), which can achieve basic document retrieval and simple Q&A, but still have significant technical defects:

[0004] Insufficient multi-source data processing capability: only text data can be processed, image and sensor data cannot be integrated, and semantic associations between data are not mined, resulting in low data value utilization;

[0005] Low semantic understanding accuracy: lack of support from power field knowledge graphs, fuzzy understanding of professional terms (such as "GIS device gas chamber pressure anomaly") and complex business logic (such as project acceptance process), prone to answering irrelevant questions;

[0006] Poor dialog flexibility: unable to adjust response strategies based on user's question history and context, prone to losing key information in multi-round dialog, and difficult to meet individual needs;

[0007] Lack of model optimization mechanism: no real-time feedback and parameter adjustment mechanism, unable to adapt to dynamic changes in power projects, and semantic understanding accuracy decreases over time.

[0008] In summary, existing systems cannot support the intelligent knowledge acquisition needs of power project full-process, and there is an urgent need for an intelligent Q&A solution that can integrate multi-source data, accurately analyze semantics, flexibly interact, and be continuously optimized. SUMMARY

[0009] To make up for the shortcomings of the prior art and solve the above technical problems, the application provides a power project Q&A system and method based on multi-data fusion and semantic analysis.

[0010] The application belongs to the technical field of power informatization and artificial intelligence, and particularly relates to intelligent question and answer technology in power project whole-process management, which is suitable for integrating multi-source heterogeneous data such as text, image and sensor in the whole life cycle of a power project, providing professional knowledge support for power project personnel through semantic analysis and intelligent interaction, and improving project decision efficiency and scientificity.

[0011] The technical solution adopted by the application to solve the technical problems is that the application proposes a power project question and answer system based on multi-data fusion and semantic analysis, which solves the technical problems of insufficient multi-source data processing, inaccurate semantic understanding, rigid dialogue and missing model optimization of existing power question and answer systems through data flow and instruction interaction collaborative work, including:

[0012] The multi-source data access module is used for accessing text data, image data and sensor data of the whole process of the power project, and configuring corresponding data access interfaces; the text interface is connected with the document library of the project management system, the image interface is connected with the storage of the inspection terminal, and the sensor interface is connected with the equipment monitoring platform; the data is classified according to the data type (the text includes design scheme / operation and maintenance record, the image includes inspection photo / construction video, and the sensor data includes equipment parameter / environment data), and the classified original data is temporarily stored in the temporary database, and the data source and collection time are marked;

[0013] The data cleaning and standardization module is used for calling original data from the temporary database of the multi-source data access module, and performing special cleaning for different types of data; the text is cleaned by removing special symbols and correcting wrong characters by using regular expressions, the image is cleaned by removing interference and unifying resolution by using a denoising algorithm, and the sensor data is cleaned by removing out-of-range data by using an outlier detection method; then, standardization processing is performed, that is, the text is converted into UTF-8 code, the image is converted into JPEG format and labeled with project / equipment, and the sensor data is arranged into a "time stamp-equipment number-parameter-value" structured data; finally, the standardized data is stored in the core database and an index is established;

[0014] The data fusion and association module is used to extract standardized data from the core database, pre-process the data features, i.e., convert text into a text feature vector, convert an image into a visual feature vector, and convert sensor data into a numerical feature vector, and construct a data association model based on a graph convolutional neural network (GCN), wherein the nodes are divided into three levels, the first level is a power project subject, the second level is a project associated entity, and the third level is an entity attribute / parameter, the edges are set as four types of "project contains entity", "entity associated document", "entity corresponding sensor" and "sensor output parameter" according to the entity association relationship and are given initial weights, the model is trained through two layers of graph convolution layers, the first layer calculates the contribution of adjacent node features, and the second layer strengthens the association through nonlinear transformation, the loss target is iteratively trained according to the matching of node feature similarity and business association degree, and the iteration is performed for 50 rounds until the loss value is less than 0.05, and finally a unified data representation of a three-layer fusion vector of "project-entity-parameter" is generated and stored in a fusion database and a node association index table is established.

[0015] The loss target refers to a core optimization direction for measuring the difference between the model output result and the expected target in the graph convolutional neural network (GCN) model training process, and the essence is to guide the model parameter adjustment by defining the matching degree of the model prediction result and the actual business logic, so as to ensure that the multi-source data association relationship learned by the model conforms to the actual business rules of the power project.

[0016] The edges according to the entity association relationship refer to the "edges" connecting different levels of nodes (the first level of "power project subject", the second level of "project associated entity", and the third level of "entity attribute / parameter") in the graph convolutional neural network (GCN) model constructed by the data fusion and association module, and the type division basis is the actual business association logic between different entities in the power project; for example, the association logic between the "220kV substation construction project" (the first level node) and the "No. 3 main transformer" (the second level node) is "project contains entity", so the edge between them is defined as the "project contains entity" type; the association logic between the "No. 3 main transformer" (the second level node) and the "main transformer operation and maintenance record" (the second level node) is "entity associated document", so the edge between them is defined as the "entity associated document" type, and so on.

[0017] Semantic parsing and model training module: Used to call the unified data representation of the fused database, import a power industry knowledge graph (RDF format storage, including subject-verb-object triples and industry authority annotations) containing five knowledge dimensions (power equipment knowledge, project management process, professional terminology definition, fault handling solution, and data association rules), and perform training in conjunction with the open-source large language model Llama2-7B: First, pre-training is performed through self-supervised learning (dual tasks: 15% professional terminology mask prediction and "question-answer" matching judgment, using 500,000 textualized data entries from the fused data set + 100,000 knowledge graph question-and-answer data entries). The training process involved 30 rounds of practice with a learning rate of 2e-5. The model was stopped when the MLM accuracy on the validation set was ≥85% and the question-answer matching accuracy was ≥90%. Then, fine-tuning was performed using transfer learning (using 10,000 historical question-answer logs, 5,000 scenario-simulated question-answers, and 3,000 error correction data points; freezing the first 10 layers, the last 5 layers after training, and the knowledge graph embedding module; iterating for 20 rounds with a cosine semantic similarity ≥0.85). Finally, 1,000 professional questions were used for testing (covering equipment malfunctions, process consultations, parameter interpretation, and solution recommendations) to ensure semantic understanding accuracy ≥90%, answer accuracy ≥88%, and professional terminology recognition rate ≥95%.

[0018] Multi-turn dialogue interaction module: This module receives text / voice questions input by users via PC / mobile (voice is automatically converted to text). It first preprocesses the questions (removing redundant characters and breaking down complex questions), then calls the parsing results (including core entity IDs, question intent, and contextual information) output by the semantic parsing and model training module. Based on deep reinforcement learning (DRL), it dynamically adjusts the response strategy—the agent is a dialogue decision model, the environment is a fusion database + dialogue log library, and the action space includes three categories: "directly generate answers," "supplement data query," and "follow up with the user." The reward function = basic reward (1 point for covering the requirements) + contextual reward (0.5 points for using historical information) + efficiency reward (0.3 points for each round of generation, decreasing by 0.1 points for each additional round). If information is missing, follow up with questions "core information first, then details." After the answer is complete, it is structured according to "anomaly type - time - measure - source" (with annotations for technical terms), and the presentation format is adjusted according to user habits. Finally, the answer is pushed to the interactive interface, and dialogue data (including user ID, question time, parsing results, feedback, and dialogue round) is simultaneously recorded in the dialogue log library.

[0019] The real-time feedback optimization module extracts data from the dialogue log database hourly, categorizing user feedback into three types: positive (satisfied), negative (unsatisfied), and neutral (needs supplementation). Negative feedback is automatically labeled with the problem type (semantic parsing error, missing data association, chaotic answer format, insufficient terminology explanation; if the cause is unclear, manual labeling occurs within 24 hours). Based on the labeling, optimization instructions are generated (e.g., adjusting terminology recognition weights, supplementing graph neural network edge associations, optimizing answer templates), and distributed to the corresponding modules for parameter adjustments (the semantic module modifies the terminology embedding layer weights and conducts small-batch testing; the fusion module updates the graph neural network edge structure and performs 10 rounds of retraining). A weekly evaluation is performed every Sunday at 23:00 (indicators: semantic understanding accuracy, answer accuracy, user satisfaction, and average dialogue rounds). If any indicator drops by more than 5%, the optimization instructions are urgently rolled back; otherwise, the optimization focus for the next cycle is determined to ensure that semantic understanding accuracy remains ≥90% and user satisfaction ≥85% in the long term.

[0020] Furthermore, the temporary database of the multi-source data access module has built-in data classification tags, including "Project Design Phase - Text - XX Scheme", "Equipment Inspection - Image - XX Equipment Photo", and "Main Transformer Monitoring - Sensor - Current Parameter", and supports quick retrieval by data source.

[0021] Furthermore, the outlier detection algorithm of the data cleaning and standardization module adopts the 3σ principle, which automatically marks and removes values ​​in the sensor data that exceed the range of "mean ± 3 times standard deviation", such as data where the current parameter suddenly spikes to more than 10 times the normal range.

[0022] Furthermore, in the graph neural network of the data fusion and association module, the initial weight of the "entity-associated document" edge is set to 0.6, the initial weight of the "sensor output parameter" edge is set to 0.8, and after training, the feature similarity of the "device entity-corresponding parameter" node needs to be ≥0.9, and the feature similarity of the "non-associated entity-data" node needs to be ≤0.3.

[0023] Furthermore, in the knowledge graph of the semantic parsing and model training module, the knowledge authority level specified by the national standard is marked as 1.0, and the internal enterprise standard is marked as 0.8. It also supports quick retrieval of triples by "equipment type" and "project stage".

[0024] Furthermore, in the question intent classification of the multi-turn dialogue interaction module, the "query" intent includes the subcategories of "operation and maintenance anomaly query" and "parameter query", and the "fault" intent includes the subcategories of "fault cause analysis" and "fault handling solution". The intent prediction probability must be ≥0.92 to determine the final intent.

[0025] Further, for the small-batch testing of the real-time feedback optimization module, 100 questions containing the target terms are randomly selected. If the term mis-identification rate drops below 1%, it is confirmed that the parameter adjustment is effective; for the weekly evaluation dataset, duplicate data needs to be removed, and the data volume ≥ 500 items.

[0026] A power item energy Q&A method based on multi-data fusion and semantic parsing includes the following steps:

[0027] Step 1, multi-source data access and preliminary classification: Start the multi-source data access module, configure the text / image / sensor interfaces to dock with the corresponding systems, collect power project data in real time, classify it by type (text / image / sensor), mark the source (such as "transmission line inspection terminal") and collection time, and temporarily store it in the temporary database;

[0028] Step 2, data cleaning and standardization processing: Call the original data in the temporary database, use regular expressions to remove special symbols from the text and correct "electric cable" to "cable", use Gaussian denoising and anti-reflection for images and unify the resolution to 1920×1080, and use the 3σ principle to remove outliers from sensor data; then standardize (UTF-8 encoding for text, JPEG format + project / device tags for images, and structured sensor data), store it in the core database, and create indexes according to project nodes / device types;

[0029] Step 3, data fusion and association based on GCN: Extract the standardized data from the core database, convert text word embeddings into feature vectors, use CNN to extract visual vectors for images, and normalize sensor data into numerical vectors; construct three-level nodes (first-level project, second-level entity, third-level attribute) and four types of edges (with initial weights); use a two-layer GCN training model (the first layer calculates the feature contributions of adjacent nodes, and the second layer performs non-linear transformation), iterate 50 rounds until the loss value < 0.05, generate a "project-entity-parameter" fusion vector, and store it in the fusion database;

[0030] Step 4, semantic parsing model training and optimization: Import the knowledge graph in the power field, perform self-supervised pre-training using Llama2-7B (dual tasks + 500,000 + 100,000 data items, 30 rounds), then perform transfer fine-tuning using 10,000 + 5,000 + 3,000 data items (freeze the first 10 layers), and finally test and meet the standard with 1,000 questions;

[0031] Step 5, multi-round dialogue interaction with reinforcement learning: Receive the user's question and preprocess it, call the semantic parsing result, query data / generate answers / prompt based on the DRL strategy, push the structured answer, and record the dialogue data in the log library;

[0032] Step 6, real-time feedback and model optimization: Extract and classify the log data every hour, generate optimization instructions to adjust the module parameters; evaluate the indicators weekly, if the reduction exceeds 5%, perform emergency optimization, otherwise formulate the key points for the next cycle.

[0033] Furthermore, the text data mentioned in step 1 includes the design scheme of the 220kV substation construction project, GIS equipment ledger, and main transformer operation and maintenance records, while the sensor data includes the voltage / current parameters of the No. 3 main transformer and the temperature and humidity data of the construction site.

[0034] Furthermore, in step 2, the text data is renamed according to "Project Stage - Data Type - File Name", such as "Construction Stage - Text - No. 3 Main Transformer Installation and Acceptance Report"; the sensor data is organized according to the format "2024-05-2014:30-No. 3 Main Transformer-Oil Temperature-65℃".

[0035] Furthermore, in step 3, the feature calculation logic of the first graph convolutional layer is "current node feature = adjacent node feature × sum of edge weights + current node initial feature". The edge weights are dynamically adjusted according to the association type. For example, the edge weights of "sensor output parameters" are increased by 10%-20% according to the importance of the parameters.

[0036] Furthermore, if the test fails to meet the target in step 4, the pre-training dataset is adjusted (e.g., the proportion of "fault handling solution" data is increased to 30%) or supplemented with fine-tuned data (e.g., 1000 question-answer pairs are added for the "project acceptance" scenario), and retrained until the target is met.

[0037] Furthermore, in step 5, follow-up questions include clarification questions (such as "Is 'recent' limited to the last 1 / 3 / 6 months?") and recommendation questions (such as "No handling measures were found, are you recommending a standard solution for the same type of anomaly?"), and the number of follow-up questions shall not exceed 2 to ensure interaction efficiency.

[0038] Furthermore, the calculation method for the weekly evaluation indicators mentioned in step 6 is as follows:

[0039] Semantic understanding accuracy = (Number of questions that correctly identified intent and entity / Total number of questions) × 100%;

[0040] Answer accuracy rate = (Number of answers matching the standard answer / Total number of answers) × 100%;

[0041] User satisfaction = (Number of positive feedback / Total number of feedback) × 100%;

[0042] Average number of dialogue rounds = Total number of dialogue rounds / Total number of dialogues.

[0043] Furthermore, if the misidentification rate of "GIS equipment" reaches 5% in step 6, the optimization instruction is to "lower the semantic matching threshold of 'GIS equipment' and its sub-concepts in the knowledge graph from 0.85 to 0.8, and at the same time increase the binding weight of 'GIS equipment' with 'gas chamber pressure' and 'SF6 gas' by 20%".

[0044] The beneficial effects of this invention are as follows:

[0045] 1. The power project question answering system and method based on multi-data fusion and semantic parsing described in this invention uses graph convolutional neural networks (GCN) to perform feature preprocessing and association modeling on multi-source heterogeneous data of power projects, including text, images, and sensors (constructing three-level nodes, four types of association edges, and iterative training), to achieve deep semantic association and unified data representation of multi-source data, and solves the technical problems of insufficient multi-source data processing capabilities and low data value utilization rate in existing systems;

[0046] 2. The power-related question-answering system and method based on multi-data fusion and semantic parsing described in this invention uses a knowledge graph in the power field (including five knowledge dimensions) combined with the open-source large language model Llama2-7B for self-supervised pre-training (dual-task training) and transfer learning fine-tuning (freezing the bottom layer to train the top layer) to achieve accurate semantic parsing of power professional terms and complex business logic. This solves the technical problems of low semantic understanding accuracy and easy to give irrelevant answers in existing systems.

[0047] 3. The power project question-answering system and method based on multi-data fusion and semantic parsing described in this invention extracts dialogue data by hourly classification and labeling through a real-time feedback optimization module, generates optimization instructions to adjust module parameters, and iteratively optimizes the model by weekly evaluation (four core indicators), so as to realize the continuous adaptation of model performance to the dynamic changes of power projects. This solves the technical problems of the lack of model optimization mechanism in existing systems and the decline in semantic understanding accuracy after long-term use. Attached Figure Description

[0048] The invention will now be further described with reference to the accompanying drawings.

[0049] Figure 1 This is a flowchart illustrating the collaborative process of each module in this invention;

[0050] Figure 2 This is a flowchart of the multi-source data access and preliminary classification process of this invention;

[0051] Figure 3 This is a flowchart of the data cleaning and standardization process of this invention;

[0052] Figure 4 This is a flowchart of the data fusion and association process based on graph neural networks in this invention;

[0053] Figure 5 This is a flowchart of the semantic parsing model training and optimization process of this invention;

[0054] Figure 6 This is a flowchart of the multi-turn dialogue interaction based on reinforcement learning in this invention;

[0055] Figure 7 This is a flowchart of the real-time feedback and continuous model optimization process of this invention. Detailed Implementation

[0056] The technical solutions of the present invention will be clearly and completely described below with reference to the embodiments and accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0057] like Figures 1-7 As shown, this invention proposes a power-related question-and-answer system based on multi-data fusion and semantic parsing. It addresses the technical problems of insufficient multi-source data processing, inaccurate semantic understanding, rigid dialogue, and lack of model optimization in existing power-related question-and-answer systems through collaborative data flow and command interaction. The system includes:

[0058] Multi-source data access module: Used to access text data, image data, and sensor data throughout the entire power project process, and configured with corresponding data access interfaces; the text interface connects to the project management system document library, the image interface connects to the inspection terminal storage, and the sensor interface connects to the equipment monitoring platform. It performs preliminary classification according to data type (text includes design schemes / operation and maintenance records, images include inspection photos / construction images, and sensor data includes equipment parameters / environmental data), and temporarily stores the classified raw data in a temporary database, marking the data source and collection time;

[0059] The data cleaning and standardization module is used to retrieve raw data from the temporary database of the multi-source data access module and perform special cleaning for different types of data. Text is cleaned by removing special symbols / correcting typos using regular expressions, images are cleaned by noise reduction algorithms to eliminate interference / unify resolution, and sensor data is cleaned by outlier detection to remove out-of-range data. Then, standardization processing is performed, namely, text is converted to UTF-8 encoding, images are converted to JPEG format and project / device tags are added, and sensor data is organized into structured data of "timestamp-device number-parameter-value". Finally, the standardized data is stored in the core database and an index is created.

[0060] The data fusion and association module is used to extract standardized data from the core database, perform feature preprocessing on the data (i.e., convert text to text feature vectors, extract visual feature vectors from images, and convert sensor data to numerical feature vectors), and construct a data association model based on a graph convolutional neural network (GCN). Nodes are divided into three levels: the first level represents the main body of the power project, the second level represents the project-related entities, and the third level represents entity attributes / parameters. Edges are categorized into four types based on entity association relationships: "project-included entities," "entity-related documents," "entity-corresponding sensors," and "sensor output parameters," and assigned initial weights. The model is trained through two layers of graph convolutional layers. The first layer calculates the feature contribution of adjacent nodes, and the second layer uses nonlinear transformation to strengthen the association. Iterative training is performed with "node feature similarity matching with business relevance" as the loss objective, iterating 50 times until the loss value is <0.05. Finally, a unified data representation of the three-layer fusion vector of "project-entity-parameter" is generated, stored in the fusion database, and a node association index table is established.

[0061] The "edge-based entity association relationship" refers to the "edge" used to connect nodes at different levels (Level 1 "Power Project Entity", Level 2 "Project Related Entities", Level 3 "Entity Attributes / Parameters") in the Graph Convolutional Neural Network (GCN) model constructed by the data fusion and association module. Its type classification is based on the actual business association logic between different entities in the power project. For example, the association logic between "220kV Substation Construction Project" (Level 1 node) and "No. 3 Main Transformer" (Level 2 node) is "Project Contains Entities", so the edge between them is defined as the "Project Contains Entities" type; the association logic between "No. 3 Main Transformer" (Level 2 node) and "Main Transformer Operation and Maintenance Record" (Level 2 node) is "Entity Related Document", so the edge between them is defined as the "Entity Related Document" type, and so on.

[0062] Semantic parsing and model training module: Used to call the unified data representation of the fused database, import a power industry knowledge graph (RDF format storage, including subject-verb-object triples and industry authority annotations) containing five knowledge dimensions (power equipment knowledge, project management process, professional terminology definition, fault handling solution, and data association rules), and perform training in conjunction with the open-source large language model Llama2-7B: First, pre-training is performed through self-supervised learning (dual tasks: 15% professional terminology mask prediction and "question-answer" matching judgment, using 500,000 textualized data entries from the fused data set + 100,000 knowledge graph question-and-answer data entries). The training process involved 30 rounds of practice with a learning rate of 2e-5. The model was stopped when the MLM accuracy on the validation set was ≥85% and the question-answer matching accuracy was ≥90%. Then, fine-tuning was performed using transfer learning (using 10,000 historical question-answer logs, 5,000 scenario-simulated question-answers, and 3,000 error correction data points; freezing the first 10 layers, the last 5 layers after training, and the knowledge graph embedding module; iterating for 20 rounds with a cosine semantic similarity ≥0.85). Finally, 1,000 professional questions were used for testing (covering equipment malfunctions, process consultations, parameter interpretation, and solution recommendations) to ensure semantic understanding accuracy ≥90%, answer accuracy ≥88%, and professional terminology recognition rate ≥95%.

[0063] Multi-turn dialogue interaction module: This module receives text / voice questions input by users via PC / mobile (voice is automatically converted to text). It first preprocesses the questions (removing redundant characters and breaking down complex questions), then calls the parsing results (including core entity IDs, question intent, and contextual information) output by the semantic parsing and model training module. Based on deep reinforcement learning (DRL), it dynamically adjusts the response strategy—the agent is a dialogue decision model, the environment is a fusion database + dialogue log library, and the action space includes three categories: "directly generate answers," "supplement data query," and "follow up with the user." The reward function = basic reward (1 point for covering the requirements) + contextual reward (0.5 points for using historical information) + efficiency reward (0.3 points for each round of generation, decreasing by 0.1 points for each additional round). If information is missing, follow up with questions "core information first, then details." After the answer is complete, it is structured according to "anomaly type - time - measure - source" (with annotations for technical terms), and the presentation format is adjusted according to user habits. Finally, the answer is pushed to the interactive interface, and dialogue data (including user ID, question time, parsing result, feedback, and dialogue round) is simultaneously recorded in the dialogue log library.

[0064] The real-time feedback optimization module extracts data from the dialogue log database hourly, categorizing user feedback into three types: positive (satisfied), negative (unsatisfied), and neutral (needs supplementation). Negative feedback is automatically labeled with the problem type (semantic parsing error, missing data association, chaotic answer format, insufficient terminology explanation; if the cause is unclear, manual labeling occurs within 24 hours). Based on the labeling, optimization instructions are generated (e.g., adjusting terminology recognition weights, supplementing graph neural network edge associations, optimizing answer templates), and distributed to the corresponding modules for parameter adjustments (the semantic module modifies the terminology embedding layer weights and conducts small-batch testing; the fusion module updates the graph neural network edge structure and performs 10 rounds of retraining). A weekly evaluation is performed every Sunday at 23:00 (indicators: semantic understanding accuracy, answer accuracy, user satisfaction, and average dialogue rounds). If any indicator drops by more than 5%, the optimization instructions are urgently rolled back; otherwise, the optimization focus for the next cycle is determined to ensure that semantic understanding accuracy remains ≥90% and user satisfaction ≥85% in the long term.

[0065] As a specific embodiment of the present invention, the temporary database of the multi-source data access module has built-in data classification tags, including "Project Design Phase - Text - XX Scheme", "Equipment Inspection - Image - XX Equipment Photo", "Main Transformer Monitoring - Sensor - Current Parameter", and supports quick retrieval by data source.

[0066] As a specific embodiment of the present invention, the outlier detection algorithm of the data cleaning and standardization module adopts the 3σ principle, which automatically marks and removes values ​​in the sensor data that exceed the range of "mean ± 3 times standard deviation", such as data where the current parameter suddenly spikes to more than 10 times the normal range.

[0067] As a specific embodiment of the present invention, in the graph neural network of the data fusion and association module, the initial weight of the "entity-associated document" edge is set to 0.6, the initial weight of the "sensor output parameter" edge is set to 0.8, and after training, the feature similarity of the "device entity-corresponding parameter" node needs to be ≥0.9, and the feature similarity of the "non-associated entity-data" needs to be ≤0.3.

[0068] As a specific embodiment of the present invention, the knowledge graph of the semantic parsing and model training module is labeled with a knowledge authority level of 1.0 as stipulated by national standards and 0.8 as stipulated by enterprise internal standards, and supports quick retrieval of triples by "equipment type" and "project stage".

[0069] As a specific embodiment of the present invention, in the question intent classification of the multi-turn dialogue interaction module, the "query" intent includes the subcategories of "operation and maintenance anomaly query" and "parameter query", and the "fault" intent includes the subcategories of "fault cause analysis" and "fault handling solution". The intent prediction probability must be ≥0.92 to determine the final intent.

[0070] As a specific implementation manner of the present invention, for the small-batch test of the real-time feedback optimization module, 100 problems containing target terms are randomly selected. If the term misrecognition rate drops below 1%, it is confirmed that the parameter adjustment is effective; for the weekly evaluation dataset, duplicate data needs to be removed, and the data volume ≥ 500 items.

[0071] A power item energy Q&A method based on multi-data fusion and semantic parsing includes the following steps:

[0072] Step 1, multi-source data access and preliminary classification: Start the multi-source data access module, configure text / image / sensor interfaces to dock with the corresponding systems, collect power project data in real time, classify it by type (text / image / sensor), mark the source (such as "transmission line inspection terminal") and the collection time, and temporarily store it in the temporary database;

[0073] Step 2, data cleaning and standardization processing: Call the original data in the temporary database. For text, use regular expressions to remove special symbols and correct "electric cable" to "cable". For images, use Gaussian denoising to eliminate reflections and unify the resolution to 1920×1080. For sensor data, use the 3σ principle to remove outliers; then standardize (text UTF-8 encoding, image JPEG format + project / device tags, sensor data structured), store it in the core database, and build indexes according to project nodes / device types;

[0074] Step 3, data fusion and association based on GCN: Extract the standardized data in the core database. Convert text using word embedding into feature vectors, extract visual vectors from images using CNN, and normalize sensor data into numerical vectors; Construct three-level nodes (first-level project, second-level entity, third-level attribute) and four types of edges (with initial weights); Use a two-layer GCN training model (the first layer calculates the feature contributions of adjacent nodes, and the second layer performs non-linear transformation), iterate 50 rounds until the loss value < 0.05, generate a "project-entity-parameter" fusion vector, and store it in the fusion database;

[0075] Step 4, semantic parsing model training and optimization: Import the knowledge graph in the power field, perform self-supervised pre-training using Llama2-7B (dual tasks + 500,000 + 100,000 data items, 30 rounds), then perform transfer fine-tuning using 10,000 + 5,000 + 3,000 data items (freeze the first 10 layers), and finally test with 1,000 problems to meet the standards;

[0076] Step 5, multi-round dialogue interaction based on reinforcement learning: Receive the user's question and preprocess it, call the semantic parsing result, query data / generate answers / prompt based on the DRL strategy, push the structured answer, and record the dialogue data in the log library;

[0077] Step 6, real-time feedback and model optimization: Extract and classify the log data every hour to generate optimization instructions to adjust the module parameters; Evaluate the indicators weekly. If the reduction exceeds 5%, perform emergency optimization, otherwise formulate the key points for the next cycle.

[0078] As a specific embodiment of the present invention, the text data in step 1 includes the design scheme of the 220kV substation construction project, the GIS equipment ledger, and the main transformer operation and maintenance records, and the sensor data includes the voltage / current parameters of the No. 3 main transformer and the temperature and humidity data of the construction site.

[0079] As a specific embodiment of the present invention, in step 2, the text data is renamed according to "project stage-data type-file name", such as "construction stage-text-No. 3 main transformer installation and acceptance report"; the sensor data is organized according to the format "2024-05-2014:30-No. 3 main transformer-oil temperature-65℃".

[0080] In one specific embodiment of the present invention, the feature calculation logic of the first graph convolutional layer in step 3 is "current node feature = adjacent node feature × sum of edge weights + current node initial feature". The edge weights are dynamically adjusted according to the association type. For example, the edge weights of "sensor output parameters" are increased by 10%-20% according to the importance of the parameters.

[0081] As a specific implementation of the present invention, if the test fails to meet the standard in step 4, the pre-training dataset is adjusted (e.g., the proportion of "fault handling solution" data is increased to 30%) or supplemented with fine-tuned data (e.g., 1000 question-answer pairs of the "project acceptance" scenario are added) and retrained until the indicator meets the standard.

[0082] As a specific embodiment of the present invention, the follow-up questions in step 5 include clarification questions (such as "Is 'recent' limited to the last 1 / 3 / 6 months?") and recommendation questions (such as "No handling measures were found, is a standard solution for the same type of abnormality recommended?"), and the number of follow-up questions does not exceed 2 to ensure interaction efficiency.

[0083] As a specific embodiment of the present invention, the calculation method of the weekly evaluation index in step 6 is as follows:

[0084] Semantic understanding accuracy = (Number of questions that correctly identified intent and entity / Total number of questions) × 100%;

[0085] Answer accuracy rate = (Number of answers matching the standard answer / Total number of answers) × 100%;

[0086] User satisfaction = (Number of positive feedback / Total number of feedback) × 100%;

[0087] Average number of dialogue rounds = Total number of dialogue rounds / Total number of dialogues.

[0088] As a specific implementation manner of the present invention, if the misrecognition rate of the "GIS device" in step 6 reaches 5%, the optimization instruction is "lower the semantic matching threshold of the 'GIS device' and its sub - concepts in the knowledge graph from 0.85 to 0.8, and at the same time increase the binding weight of the 'GIS device' with the 'gas chamber pressure' and 'SF6 gas' by 20%".

[0089] As Figures 1-7 shown, to make the technical solution of the present invention clearer, the following combines the example of the "220kV substation construction project" to detail the specific implementation process of the present invention:

[0090] Multi - source data access and preliminary classification:

[0091] Start the multi - source data access module and configure three types of interfaces: the text interface is docked with the "document library of the 220kV substation project management system", the image interface is docked with the "transmission line inspection terminal storage", and the sensor interface is docked with the "No. 3 main transformer monitoring platform";

[0092] The interfaces collect data in real - time: text data (design plan and operation and maintenance records of the No. 3 main transformer), image data (inspection photos of the No. 3 main transformer and construction site images), sensor data (oil temperature / current parameters of the No. 3 main transformer, ambient temperature and humidity of the substation);

[0093] After classification by type, temporarily store it in the temporary database, and mark the data source (such as "No. 3 main transformer monitoring sensor") and the collection time (such as "2024 - 05 - 20 14:30").

[0094] Data cleaning and standardization processing:

[0095] Special cleaning: Use regular expressions to remove special symbols such as "#" and "@" from the text data, and correct "electric cable" to "cable"; Use the Gaussian denoising algorithm to eliminate the reflection of the inspection photos in the image data, and unify the resolution to 1920×1080; Use the 3σ principle to剔除 the abnormal value of "the current of the No. 3 main transformer suddenly soars to 10 times the normal range" from the sensor data;

[0096] Standardization: Convert the text data to UTF - 8 encoding and rename it to "construction stage - text - operation and maintenance records of the No. 3 main transformer"; Convert the image data to JPEG format and add the label "220kV substation project - No. 3 main transformer"; Organize the sensor data as "2024 - 05 - 20 14:30 - No. 3 main transformer - oil temperature - 65°C";

[0097] Store it in the core database and establish an index according to "project stage (construction) - equipment type (main transformer)".

[0098] Data fusion and association based on GCN:

[0099] Note: In the original text, there is a Chinese character "剔除" which is not recognized in the English translation software. I have used "剔除" to represent the action of removing outliers in the sensor data for the purpose of translation. If there is a more accurate English expression for this operation, it can be adjusted accordingly.Feature preprocessing: Text data (maintenance records of transformer No. 3) was converted into text feature vectors using word embedding technology; image data (inspection photos of transformer No. 3) was extracted into visual feature vectors using convolutional neural networks; sensor data (oil temperature parameters) was converted into numerical feature vectors using feature normalization.

[0100] Construct the GCN model: Nodes are set as "220kV substation project" (Level 1), "No. 3 main transformer" (Level 2), and "main transformer oil temperature parameters" (Level 3); edges are set as "project-included entities" (Level 1-Level 2, weight 0.7), "entity-corresponding sensors" (Level 2-sensor nodes, weight 0.8), and "sensor output parameters" (sensor nodes-Level 3, weight 0.9);

[0101] Training the model: Input feature vectors and edge weights, train through two layers of GCN for 50 rounds until the loss value is <0.05 (e.g., the feature similarity between "No. 3 main transformer" and "main transformer oil temperature parameter" is ≥0.9); generate the fusion vector of "220kV substation project - No. 3 main transformer - main transformer oil temperature parameter" and store it in the fusion database.

[0102] Semantic parsing model training and optimization:

[0103] Imported knowledge graph of the power industry: including ternary pairs such as "No. 3 main transformer - rated oil temperature - 60℃" and "GIS equipment - includes - air chamber pressure sensor", with national standard knowledge authority rating of 1.0;

[0104] Self-supervised pre-training: Using 500,000 fused textual data entries (e.g., "No. 3 main transformer oil temperature sensor output 65℃ on 2024-05-20") + 100,000 knowledge graph question-and-answer data entries (e.g., question "What is the rated oil temperature of No. 3 main transformer?", answer "60℃"), training for 30 rounds, the validation set MLM accuracy was 86% and the question-and-answer matching accuracy was 91%, then pre-training was stopped;

[0105] Migration fine-tuning: Using 10,000 historical question-and-answer logs (such as "Installation and acceptance standards for main transformer No. 3") + 5,000 scenario simulation questions and answers (1,250 during the construction phase) + 3,000 error correction data, freeze the first 10 layers of the model, train the last 5 layers and the knowledge graph embedding module, iterate 20 times, and achieve a semantic similarity ≥ 0.85.

[0106] Test: Using 1000 professional questions (250 equipment fault queries, 250 process consultations, etc.), the semantic understanding accuracy rate was 92%, the answer accuracy rate was 89%, and the professional terminology recognition rate was 96%, meeting the standards for activation.

[0107] Multi-turn dialogue interaction based on reinforcement learning:

[0108] Users input the following question via PC: "What anomalies are there in the recent maintenance records of main transformer No. 3? Please explain the time of the anomaly and the handling measures." The module preprocesses the question by removing redundant characters such as "please" and breaking it down into "anomaly type," "anomaly time," and "handling measures."

[0109] Semantic parsing: Identify the core entities "No. 3 main transformer" and "Operation and maintenance record", with the intent being "query type", and associate it with the historical query "No. 3 main transformer rated oil temperature" (association degree 0.8) to generate the parsing result;

[0110] Reinforcement learning response: The agent queries the fusion database and extracts 5 maintenance anomalies of the No. 3 main transformer in the past 3 months. The information is complete (including time and measures). If the agent selects "generate answer directly", the total reward = 1 (basic) + 0.5 (context) + 0.3 (1 round) = 1.8.

[0111] Answer optimization: Organize by "Abnormality type-time-handling measures-source" (e.g., "2024-04-10 10:20, oil temperature too high 72℃, handling measures: reduce load and check cooling system, source: No. 3 main transformer operation and maintenance record"), and annotate "oil temperature too high" as "refers to exceeding the rated oil temperature by 60℃, which may cause equipment overheating";

[0112] Push the answer to the PC and record the dialogue data (user ID, question time, analysis result, feedback "satisfied", dialogue round 1) to the log database.

[0113] Real-time feedback and continuous model optimization:

[0114] Log data is extracted every hour: categorized as positive feedback (“satisfactory”) or negative feedback (e.g., “GIS equipment misidentified as transformer”); negative feedback is labeled “semantic parsing error”.

[0115] Optimization instructions are generated: For the "5% false recognition rate of GIS equipment", the instructions are "lower the semantic matching threshold of 'GIS equipment' in the knowledge graph from 0.85 to 0.8, and increase the binding weight with 'gas chamber pressure' by 20%";

[0116] Adjustments were made: The semantic parsing module modified the weights of the term embedding layer, and a small batch test of 100 questions containing "GIS equipment" reduced the false recognition rate to 0.8%.

[0117] Weekly evaluation: On Sunday at 23:00, 500 "question-answer" pairs were evaluated. The semantic understanding accuracy rate was 93%, the answer accuracy rate was 90%, the user satisfaction rate was 87%, and the average number of dialogue rounds was 1.2. The indicators were stable, and the optimization focus for the next cycle was "improving the correlation of image data".

[0118] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of the present invention is defined by the appended claims and their equivalents.

Claims

1. A power item question and answer system based on multi-data fusion and semantic analysis, characterized in that, Comprise working cooperatively with instructions through data flow conversion and interaction: Multi-source data access module for accessing text, image and sensor data of power projects, preliminary classification and temporary storage according to data type; Data cleaning and standardization module for performing special cleaning and standardization processing on the original data, and storing the processed data into the core database; Data fusion and association module for performing feature preprocessing and association modeling on the standardized data based on graph convolutional neural network (GCN), generating unified data representation and storing; Semantic analysis and model training module for training and verifying the unified data representation by combining power domain knowledge graph and large language model, ensuring semantic understanding and answer accuracy; Multi-round dialogue interaction module for receiving user questions, calling semantic analysis results and adjusting response strategy based on reinforcement learning, pushing answers and recording dialogue data; Real-time feedback optimization module for extracting dialogue data classification labels, generating optimization instructions to adjust module parameters, regularly evaluating model performance and iteratively optimizing.

2. The power item question answering system based on multi-data fusion and semantic parsing of claim 1, wherein, The text data accessed by the multi-source data access module includes project design scheme, equipment account and operation and maintenance record, the image data includes equipment inspection photos and construction site videos, and the sensor data includes equipment operation parameters and environmental monitoring data. 3.The power item Q&A system based on multi-data fusion and semantic parsing of claim 1, wherein, The special cleaning of the data cleaning and standardization module includes: removing special symbols and correcting wrong characters in text data by regular expression, removing interference in image data by denoising algorithm, and removing out-of-range data in sensor data by outlier detection.

4. The power item Q&A system based on multi-data fusion and semantic parsing of claim 1, wherein, In the GCN model constructed by the data fusion and association module, the nodes are divided into three levels: the first level is the power project main body, the second level is the project associated entity, and the third level is the entity attribute or parameter; the edges are divided into four categories according to the entity association relationship: "project contains entity", "entity associated document", "entity corresponding sensor" and "sensor output parameter".

5. The power domain question answering system based on multi-data fusion and semantic parsing of claim 1, wherein, The power domain knowledge graph imported by the semantic analysis and model training module contains five knowledge dimensions: power equipment knowledge, project management process, professional term definition, fault handling scheme and data association rule.

6. The power domain question answering system based on multi-data fusion and semantic parsing of claim 1, wherein, The action space of the multi-round dialogue interaction module based on reinforcement learning includes three types of actions: "directly generating answers", "supplementing query data" and "asking users for information".

7. The power domain question answering system based on multi-data fusion and semantic parsing of claim 1, wherein, The user feedback is divided into three categories: positive, negative and neutral. Negative feedback is labeled with semantic analysis errors, data association missing, answer format disorder or insufficient term explanation, and corresponding optimization instructions are generated.

8. The method of claim 1-7, wherein the method is applied to the system of claim 1-7. The steps include: ​ Step 1: Access multi-source data of power projects and classify and temporarily store them according to type; Step 2: Perform cleaning and standardization processing on the original data and store them into the core database; Step 3: Model the standardized data based on GCN to generate unified data representation; Step 4: Train semantic analysis model combining knowledge graph and large language model and verify it; Step 5: Receive user questions, respond based on reinforcement learning and record dialogue data; Step 6: Extract dialogue data to generate optimization instructions, regularly evaluate and iteratively optimize the model. 9.The power item question answering method based on multi-data fusion and semantic parsing according to claim 8, characterized in that, The GCN model training in step 3 takes the matching of node feature similarity and power project actual business correlation as the loss target. 10.The power item question answering method based on multi-data fusion and semantic parsing according to claim 8, characterized in that, The periodic evaluation indicators in step 6 include semantic understanding accuracy, answer accuracy, user satisfaction, and average number of dialogue turns.

Images