A power plant safety regulation intelligent question and answer interaction method and system based on a large model

Abstract

This invention relates to the field of power plant technology, specifically to a large-model-based intelligent question-and-answer interaction method and system for power plant safety regulations. The large-model-based intelligent question-and-answer interaction method for power plant safety regulations includes the following steps: performing a first preprocessing on multi-source heterogeneous data to generate a knowledge graph and multimodal vector embeddings, constructing a multimodal knowledge base for power plant safety regulations; receiving user-inputted question-and-answer requests, performing a second preprocessing on the requests to generate standardized query data; based on a retrieval enhancement generation mechanism, recalling and optimizing related knowledge fragments from the multimodal knowledge base through a hybrid retrieval strategy; constructing a fine-tuned large-scale model for power plant safety regulations, and inputting the standardized query data and related knowledge fragments into the fine-tuned large-scale model to generate a set of candidate answers; performing compliance verification and confidence assessment on the candidate answer set, selecting the answer with the highest confidence as the final response, and feeding back the final response and its associated safety regulation data to the user.

Description

Technical Field

[0001] This invention relates to the field of power plant technology, specifically to a power plant safety regulation intelligent question-and-answer interaction method and system based on a large model. Background Technology

[0002] Power plant safety regulations are the core principles for ensuring safety in all scenarios of power plant operation, maintenance, and emergency response. Staff need to quickly obtain accurate interpretations of safety regulations, operating procedures, and case references in their daily work and emergency response to avoid safety accidents caused by violations. With the upgrading of power plant equipment and the revision of safety regulations, multimodal and fragmented safety regulation data are increasing, which places higher demands on the accuracy, timeliness, and multi-scenario adaptability of question-and-answer interaction.

[0003] Existing methods for answering questions related to power plant safety regulations have significant shortcomings. They rely heavily on manual searching of safety regulations texts or traditional search tools, which is inefficient and prone to missing key clauses. They lack multimodal interactive capabilities and cannot adapt to diverse query needs such as voice and images. Ordinary question-and-answer systems are not deeply optimized for the field of power plant safety regulations, which can easily generate "illusionary" content that conflicts with safety regulations. Furthermore, their knowledge updates are lagging behind, making it difficult to synchronize safety regulations revisions with newly added equipment operating procedures, and thus failing to provide reliable decision support for staff. Summary of the Invention

[0004] The purpose of this invention is to provide an intelligent question-and-answer interaction method for power plant safety regulations based on a large model, so as to solve the problems of low efficiency and easy omission of key clauses in the existing power plant safety regulations question-and-answer method.

[0005] To address the aforementioned problems, this invention proposes a power plant safety regulation intelligent question-and-answer interaction method based on a large model. The technical solution adopted is as follows: A power plant safety regulation intelligent question-and-answer interaction method based on a large model includes the following steps: S1: Perform the first preprocessing on the multi-source heterogeneous data of the power station to generate a power station safety regulation knowledge graph and a multimodal vector embedding with text-speech-image adaptation, and construct a power station safety regulation multimodal knowledge base; S2: Receive user input question-and-answer requests, perform second preprocessing on the question-and-answer requests, and generate standardized query data; S3: Based on the retrieval enhancement generation mechanism, a hybrid retrieval strategy is used to recall and optimize related knowledge fragments from the power plant safety regulations multimodal knowledge base; S4: Construct a large-scale model for fine-tuning in the field of power plant safety regulations, and input standardized query data and related knowledge fragments of recall and optimization into the large-scale model for fine-tuning in the field of power plant safety regulations to generate a set of candidate answers that conform to safety regulations; S5: Perform compliance verification and confidence assessment on the candidate answer set, select the answer with the highest confidence as the final response, and feed back the final response and its associated safety data to the user.

[0006] Further, in S1, the first preprocessing of the multi-source heterogeneous data of the power station to generate a power station safety regulation knowledge graph and a text-speech-image adapted multimodal vector embedding, and to construct a power station safety regulation multimodal knowledge base, includes: Collect structured power plant safety regulations and equipment operating procedures, unstructured accident cases and maintenance records, as well as multimodal equipment diagrams and operation video frames to obtain multi-source heterogeneous data; Text information from images and video frames in multi-source heterogeneous data is extracted using OCR technology, and then cleaned and normalized to obtain text data. Semantic word segmentation and named entity recognition technologies are used to annotate power plant-specific terms, and entity-relation-attribute triples are established to form a power plant safety regulation knowledge graph. Map text data and entity relationships in the power plant safety regulations knowledge graph to a unified semantic space to generate multimodal vector embeddings; By storing multimodal vector embeddings and power plant safety regulations knowledge graphs in a hybrid storage architecture consisting of a vector database and a graph database, a multimodal knowledge base for power plant safety regulations is obtained.

[0007] Further, in S2, receiving the user's input question-and-answer request and performing a second preprocessing on the request to generate standardized query data includes: The system receives user-inputted question-and-answer requests, performs modal parsing, noise filtering, and semantic enhancement on the requests, extracts core entities and scene features, and generates standardized query data. The question-and-answer requests include at least one of text, voice, and image.

[0008] Furthermore, the semantic enhancement process specifically includes: For text requests, a dedicated dictionary for the power plant domain is used to optimize word segmentation of the question-and-answer requests after modal parsing and noise filtering; core entities are identified through a pre-trained named entity recognition model, and semantic missing information is supplemented by context to transform fuzzy requests into standardized query data; For voice requests, the data is first converted into text using speech-to-text technology, and then semantic enhancement processing steps are performed for text requests to obtain standardized query data. For image requests, semantic query vectors are generated by combining image features with text descriptions to obtain standardized query data.

[0009] Furthermore, in S3, the step of retrieving relevant knowledge fragments from the power plant safety regulations multimodal knowledge base using a hybrid retrieval strategy based on the retrieval enhancement generation mechanism includes: Based on core entities, a preliminary knowledge set that matches the characteristics of the scenario is filtered out through Boolean retrieval, while irrelevant domain data is eliminated; The semantic similarity between the standardized query data and the preliminary knowledge set is calculated by vector semantic retrieval, and the top-k related knowledge fragments are recalled, where k is a positive integer greater than or equal to 2; Based on the entity relationship links of the power plant safety regulations knowledge graph, path scoring and ranking optimization are performed on the top-k knowledge fragments.

[0010] Furthermore, in S4, the construction of the fine-tuned large model in the field of power plant safety regulations includes: Construct a dataset for fine-tuning power plant safety regulations; Based on the power plant safety regulation fine-tuning dataset, LoRA lightweight fine-tuning technology is used to freeze the basic parameters of the basic large model and train the parameters of the domain adaptation layer and attention layer of the basic large model. By introducing a safety compliance loss function, the consistency between the generated answer of the large model and the safety regulations is constrained, resulting in a fine-tuned large model for power plant safety regulations. The basic large model is selected from the GPT series, ERNIE or LLaMA series pre-trained models.

[0011] Furthermore, in S5, the compliance verification and confidence assessment of the candidate answer set, and the selection of the answer with the highest confidence as the final response, includes: A compliance verification rule base was established based on the power plant safety regulations knowledge graph and the core clauses of the power plant safety regulations. By comparing the candidate answer set with the compliance verification rule base using a semantic matching algorithm, conflict detection and process compliance judgment are performed to obtain a compliant answer set; A pre-set reliability threshold is used to calculate the confidence score based on the semantic fit between the set of compliant answers and the question-and-answer request, the sufficiency of the retrieved evidence, and the matching accuracy of the safety regulations. The score is calculated using a weighted summation method. If the confidence score is greater than or equal to the confidence threshold, the highest confidence score will be selected as the final response. If the confidence score is less than the confidence threshold, the set of compliant answers needs to be supplemented or the user needs to be prompted to refine the question.

[0012] Furthermore, the intelligent question-and-answer interaction method for power plant safety regulations based on a large model also includes: Incremental updates to the power plant safety regulations multimodal knowledge base are performed, specifically as follows: Real-time collection of revised safety regulations for power plants, operating procedures for newly added equipment, and the latest accident cases yields new data. This data is then updated using a combination of incremental web crawling and manual review to address the multi-source heterogeneous data. The first preprocessing step S1 is performed on the newly added data to generate an incremental knowledge graph and vector embeddings. An incremental update algorithm is used to integrate incremental knowledge graphs and vector embeddings into a multimodal knowledge base for power plant safety regulations, thereby achieving incremental updates of the knowledge base. At the same time, redundancy cleanup and contradiction detection are carried out regularly in the knowledge base, and multi-source data conflicts are resolved based on data source priority rules.

[0013] Furthermore, the intelligent question-and-answer interaction method for power plant safety regulations based on a large model also includes: Users input follow-up questions and answers, and based on the context of historical questions and answers and the retrieved knowledge, subsequent responses are generated to achieve multi-round continuous interaction; Meanwhile, during the interaction process, user question and answer requests, response content, and operation logs are encrypted and stored, and access permissions are set for hierarchical control.

[0014] This invention also proposes a large-model-based intelligent question-and-answer interaction system for power plant safety regulations, which, based on the aforementioned large-model-based intelligent question-and-answer interaction method for power plant safety regulations, includes: The power plant safety regulations multimodal knowledge base construction module is used to perform the first preprocessing of multi-source heterogeneous data of the power plant, generate a power plant safety regulations knowledge graph and text-speech-image adapted multimodal vector embedding, and construct a power plant safety regulations multimodal knowledge base; The standardized query data generation module is used to receive user-inputted question-and-answer requests, perform a second preprocessing on the requests, and generate standardized query data. The associated knowledge fragment recall and optimization module is used to recall and optimize associated knowledge fragments from the power plant safety regulations multimodal knowledge base based on the retrieval enhancement generation mechanism and through a hybrid retrieval strategy. The candidate answer set generation module is used to build a fine-tuning large model in the field of power plant safety regulations. It inputs standardized query data and related knowledge fragments of recall and optimization into the fine-tuning large model in the field of power plant safety regulations to generate a candidate answer set that conforms to safety regulations. The final response acquisition module is used to perform compliance verification and confidence assessment on the candidate answer set, select the answer with the highest confidence as the final response, and feed the final response and its associated safety data back to the user.

[0015] Compared with the prior art, the present application has the following beneficial effects: This invention is an improved invention. It solves the pain point of low efficiency caused by traditional single text query by multimodal adaptation of text, voice and image, and supports diversified query needs. At the same time, it generates standardized query data from user input question and answer requests, which shortens the query response time. Based on domain fine-tuning of large models, staff do not need to manually flip through safety regulations texts and can quickly obtain accurate answers, adapting to the full-scenario operation needs of power plants.

[0016] The specific effects include: 1. Precise and efficient question answering: Through the technical approach of "retrieval enhancement generation mechanism + domain fine-tuning large model", the answers are ensured to originate from the authoritative multimodal knowledge base of power plant safety regulations, and are professionally and humanizedly integrated and expressed by the domain fine-tuning large model of power plant safety regulations, which significantly improves the accuracy and practicality of the answers and provides a fast response speed.

[0017] 2. Multimodal interactive features: Supports multiple input methods including text, voice, and images, adapting to complex on-site environments and greatly improving the convenience and naturalness of interaction.

[0018] 3. Strong compliance guarantee: The system introduces a compliance verification and confidence assessment process for the candidate answer set, which effectively eliminates incorrect answers that violate safety regulations and ensures the seriousness and reliability of the system.

[0019] 4. Knowledge is traceable and evolving: The answers are accompanied by complete safety regulations data, which enhances credibility.

[0020] 5. Full-scenario coverage: From the construction of a multimodal knowledge base for power plant safety regulations to Q&A interaction, it covers the entire business process of power plant operation, maintenance, and emergency response, providing a unified and accurate access point for safety regulations knowledge services for personnel in different positions.

[0021] The power plant safety regulation intelligent question-and-answer interaction method based on a large model also includes: Incremental updates to the power plant safety regulations multimodal knowledge base are performed, specifically as follows: Real-time collection of revised safety regulations for power plants, operating procedures for newly added equipment, and the latest accident cases yields new data. This data is then updated using a combination of incremental web crawling and manual review to address the multi-source heterogeneous data. The first preprocessing step S1 is performed on the newly added data to generate an incremental knowledge graph and vector embeddings. An incremental update algorithm is used to integrate incremental knowledge graphs and vector embeddings into a multimodal knowledge base for power plant safety regulations, thereby achieving incremental updates of the knowledge base. Meanwhile, the knowledge base is regularly cleaned up for redundancy and contradictions are detected, and multi-source data conflicts are resolved based on data source priority rules. An incremental update mechanism synchronizes safety regulation revisions and new procedures, improving the timeliness of knowledge, supporting multi-round interactions and visualization of complex issues, adapting to power plant equipment upgrades and safety regulation iterations, reducing overall operation and maintenance costs, and making it suitable for all scenarios of power plant operation, maintenance, and emergency response. Attached Figure Description

[0022] Figure 1 This is a schematic diagram of the intelligent question-and-answer interaction method for power plant safety regulations based on a large model, as presented in this invention. Figure 2This is a schematic diagram of the process structure for constructing and updating the multimodal knowledge base in the intelligent question-and-answer interaction method for power plant safety regulations based on a large model, as described in this invention. Figure 3 This is a schematic diagram of the process structure for large model fine-tuning and answer generation and verification in the intelligent question-and-answer interaction method for power plant safety regulations based on a large model, as described in this invention. Detailed Implementation

[0023] The embodiments of this application are described in detail below. Examples of these embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and intended to explain this application, and should not be construed as limiting this application.

[0024] It should be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of this invention are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of the invention described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover a non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0025] The following is combined with Figure 1 This application provides a detailed description of a power plant safety regulation intelligent question-and-answer interaction method based on a large model.

[0026] In this embodiment, as Figure 1 , 2 As shown in Figure 3, this invention provides a power plant safety regulation intelligent question-and-answer interaction method based on a large model, comprising the following steps: S1: Perform the first preprocessing on the multi-source heterogeneous data of the power station to generate a power station safety regulation knowledge graph and a multimodal vector embedding with text-speech-image adaptation, and construct a power station safety regulation multimodal knowledge base.

[0027] Specifically, the multi-source heterogeneous data of the power plant undergoes a first preprocessing step to generate a power plant safety regulation knowledge graph and a text-speech-image adapted multimodal vector embedding, constructing a power plant safety regulation multimodal knowledge base, including: Structured power plant safety regulations and equipment operating procedures, unstructured accident cases and maintenance records, as well as multimodal equipment diagrams and operation video frames are collected to obtain multi-source heterogeneous data. Text information from images and video frames in the multi-source heterogeneous data is extracted using OCR technology, and then cleaned and normalized to obtain text data. Semantic word segmentation and named entity recognition technologies are used to annotate power plant-specific terms, establishing entity-relationship-attribute triples to form a power plant safety regulation knowledge graph. The text data and entity relationships in the power plant safety regulation knowledge graph are mapped to a unified semantic space to generate multimodal vector embeddings. The multimodal vector embeddings and the power plant safety regulation knowledge graph are stored in a hybrid storage architecture composed of a vector database and a graph database to obtain a power plant safety regulation multimodal knowledge base. Here, cross-modal knowledge association is achieved to ensure comprehensive retrieval.

[0028] The power plant safety regulations include texts covering all scenarios of power plant operation, maintenance, and emergency response. The first preprocessing step involves cleaning, normalizing, and semantically annotating the multi-source heterogeneous data.

[0029] Vector databases use Milvus or FAISS, and graph databases use Neo4j.

[0030] S2: Receive user input for question and answer requests, perform a second preprocessing on the question and answer requests, and generate standardized query data.

[0031] Specifically, it receives user-input question-and-answer requests, performs a second preprocessing on the requests, and generates standardized query data, including: The system receives user-inputted question-and-answer requests, performs modal parsing, noise filtering, and semantic enhancement on the requests, extracts core entities and scene features, and generates standardized query data. The question-and-answer requests include at least one of text, voice, and image.

[0032] Specifically, the semantic enhancement processing involves: for text requests, using a power plant-specific dictionary to optimize word segmentation after modal parsing and noise filtering, correcting ambiguities related to homophones and technical abbreviations; identifying core entities using a pre-trained named entity recognition model, supplementing missing semantic information with context, and transforming ambiguous requests into standardized query data; for voice requests, first converting speech to text using speech-to-text technology, then performing semantic enhancement processing steps for text requests to obtain standardized query data; and for image requests, combining image features with text descriptions to generate semantic query vectors, resulting in standardized query data. This semantic enhancement processing lays the foundation for accurate retrieval.

[0033] S3: Based on the retrieval enhancement generation mechanism, a hybrid retrieval strategy is used to recall and optimize related knowledge fragments from the power plant safety regulations multimodal knowledge base.

[0034] Specifically, based on the retrieval enhancement generation mechanism, a hybrid retrieval strategy is used to recall relevant knowledge fragments from the power plant safety regulations multimodal knowledge base, including: Based on core entities, a preliminary knowledge set matching the scenario characteristics is selected through Boolean retrieval, and irrelevant domain data is eliminated. The semantic similarity between the standardized query data and the preliminary knowledge set is calculated through vector semantic retrieval, and the top-k related knowledge fragments are recalled, where k is a positive integer greater than or equal to 2. Based on the entity relationship links of the power plant safety regulations knowledge graph, the path scoring and ranking optimization of the top-k knowledge fragments are performed to further filter highly related knowledge, thereby improving the recall accuracy.

[0035] S4: Construct a large-scale model for fine-tuning in the field of power plant safety regulations, and input standardized query data and related knowledge fragments for recall and optimization into the large-scale model for fine-tuning in the field of power plant safety regulations to generate a set of candidate answers that conform to safety regulations.

[0036] Specifically, a fine-tuning large-scale model for power plant safety regulations is constructed, including: A power plant safety regulation fine-tuning dataset is constructed. Based on the power plant safety regulation fine-tuning dataset, LoRA lightweight fine-tuning technology is adopted to freeze the basic parameters of the basic large model and train only the domain adaptation layer and attention layer parameters of the basic large model to reduce training computational consumption. A safety regulation compliance loss function is introduced to constrain the consistency between the large model's generated answers and the safety regulations, avoid generating illusory content, and obtain a power plant safety regulation domain fine-tuning large model. Among them, the basic large model uses GPT series, ERNIE or LLaMA series pre-trained models.

[0037] Among them, the large-scale model for fine-tuning in the field of power plant safety regulations focuses on core semantics through an attention mechanism to generate a set of candidate answers that conform to safety regulations and are logically rigorous.

[0038] The power plant safety regulation fine-tuning dataset includes labeled data such as safety regulation Q&A pairs, clause explanations, and violation case analyses.

[0039] In one specific embodiment, a fine-tuning large model for power plant safety regulations is constructed, including: First, construct and fine-tune a dataset including safety regulation question-and-answer pairs and clause interpretations; Subsequently, based on the fine-tuned dataset, the LoRA technique was used to freeze the basic parameters of the basic large model, and only the domain adaptation layer was trained to reduce computational consumption; Next, a compliance loss function is introduced to constrain the consistency between the answer and the safety regulations. Finally, by inputting standardized query data and knowledge related to recall and optimization, a logically rigorous set of candidate answers is generated.

[0040] S5: Perform compliance verification and confidence assessment on the candidate answer set, select the answer with the highest confidence as the final response, and feed back the final response and its associated safety data to the user.

[0041] Specifically, the candidate answer set undergoes compliance verification and confidence assessment, and the answer with the highest confidence level is selected as the final response, including: Based on the power plant safety regulations knowledge graph and core clauses of the regulations, a compliance verification rule base is established. A semantic matching algorithm is used to compare the candidate answer set with the compliance verification rule base to perform conflict detection and process compliance judgment, resulting in a compliant answer set. A pre-set confidence threshold is used to calculate the confidence score based on the semantic fit between the compliant answer set and the question-and-answer request, the sufficiency of the retrieved evidence, and the matching accuracy of the safety regulations clauses. If the confidence score is greater than or equal to the confidence threshold, the highest confidence score is selected as the final response; if the confidence score is less than the confidence threshold, the compliant answer set needs to be supplemented with additional searches or the user needs to be prompted to refine the question.

[0042] This involves conducting compliance verification and confidence assessment on the candidate answer set, while eliminating answers that conflict with safety regulations or lack sufficient evidence.

[0043] Safety regulation data includes labeled answers, safety regulation clauses, reference cases, and a chain of retrieval evidence. Specifically, this includes: standardized answer text, the original text and clause number of the corresponding safety regulation clause, a summary of the associated reference case, and a visual representation of the chain of retrieval evidence. Additionally, for complex operational questions, flowcharts or equipment diagrams are output simultaneously.

[0044] The compliance verification rule base includes clause conflict detection, operational process compliance judgment, and emergency response standardization verification.

[0045] In one specific embodiment, a preset confidence threshold of 0.8 is set. If the confidence score is greater than or equal to 0.8, the confidence score is determined as the final response; if the confidence score is less than 0.8, the compliant answer set needs to be supplemented by searching or the user is prompted to refine the question.

[0046] The intelligent question-and-answer interaction method for power plant safety regulations based on a large model also includes: incrementally updating the multimodal knowledge base of power plant safety regulations, specifically: The system collects revised safety regulations for power plants, newly added equipment operating procedures, and the latest accident cases in real time to obtain new data. This new data is then updated using a combination of incremental web crawling and manual review. Step S1, the first preprocessing step, is executed on the new data to generate an incremental knowledge graph and vector embeddings. An incremental update algorithm is then used to integrate the incremental knowledge graph and vector embeddings into the multimodal knowledge base of power plant safety regulations, enabling incremental updates to the knowledge base. Data versions and update times are also marked to ensure the timeliness of the knowledge. Furthermore, the knowledge base is regularly cleaned up for redundancy and conflict detection is performed, and multi-source data conflicts are resolved based on data source priority rules.

[0047] The intelligent question-and-answer interaction method for power plant safety regulations based on large models also includes: Users input follow-up questions and answers, and based on the context of historical questions and answers and the retrieved knowledge, subsequent responses are generated to achieve multi-round continuous interaction; Meanwhile, during the interaction process, user Q&A requests, responses, and operation logs are encrypted and stored, and access permissions are controlled in a tiered manner, allowing only authorized personnel to view sensitive security data and interaction records, thus ensuring data security and compliance.

[0048] Specifically, the intelligent question-and-answer interaction method for power plant safety regulations based on a large model, as described in this invention, includes the following steps: First, users can initiate maintenance-related safety regulation consultation requests via text, voice, or images; Subsequently, the question-and-answer requests are processed through modal parsing, noise filtering, and semantic enhancement to convert them into standardized query data. Next, a hybrid retrieval strategy was used to retrieve relevant safety regulations, operating procedures, and case studies from the power plant safety regulations multimodal knowledge base; Then, the large-scale model for power plant safety regulations is fine-tuned by combining query and recall knowledge to generate a set of candidate answers that conform to safety regulations. Next, through compliance verification and confidence assessment, the best answer is selected and linked with the evidence chain to provide feedback to the user. The system also supports follow-up questions from the user, and quickly generates subsequent responses based on historical context. Finally, the knowledge base is updated incrementally in real time to synchronize the latest safety regulations and ensure the timeliness of knowledge.

[0049] This invention also provides a power plant safety regulation intelligent question-and-answer interaction system based on a large model, which, based on the above-mentioned power plant safety regulation intelligent question-and-answer interaction method based on a large model, includes: The power plant safety regulations multimodal knowledge base construction module is used to perform the first preprocessing of multi-source heterogeneous data of the power plant, generate a power plant safety regulations knowledge graph and text-speech-image adapted multimodal vector embedding, and construct a power plant safety regulations multimodal knowledge base; The standardized query data generation module is used to receive user-inputted question-and-answer requests, perform a second preprocessing on the requests, and generate standardized query data. The associated knowledge fragment recall and optimization module is used to recall and optimize associated knowledge fragments from the power plant safety regulations multimodal knowledge base based on the retrieval enhancement generation mechanism and through a hybrid retrieval strategy. The candidate answer set generation module is used to build a fine-tuning large model in the field of power plant safety regulations. It inputs standardized query data and related knowledge fragments of recall and optimization into the fine-tuning large model in the field of power plant safety regulations to generate a candidate answer set that conforms to safety regulations. The final response acquisition module is used to perform compliance verification and confidence assessment on the candidate answer set, select the answer with the highest confidence as the final response, and feed the final response and its associated safety data back to the user.

[0050] Those skilled in the art will understand that the specific operations of the above-mentioned intelligent question-and-answer interaction system for power plant safety regulations based on a large model have been referenced above. Figures 1 to 3 The method for intelligent question-and-answer interaction based on large models for power plant safety regulations is described in detail here.

[0051] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. The scope of patent protection of the present invention shall be determined by the claims. Similarly, any equivalent structural changes made based on the description and drawings of the present invention shall also be included within the scope of protection of the present invention.

Claims

1. A smart question-and-answer interaction method for power plant safety regulations based on a large model, characterized in that, Includes the following steps: S1: Perform the first preprocessing on the multi-source heterogeneous data of the power station to generate a power station safety regulation knowledge graph and a multimodal vector embedding with text-speech-image adaptation, and construct a power station safety regulation multimodal knowledge base; S2: Receive user input question-and-answer requests, perform second preprocessing on the question-and-answer requests, and generate standardized query data; S3: Based on the retrieval enhancement generation mechanism, a hybrid retrieval strategy is used to recall and optimize related knowledge fragments from the power plant safety regulations multimodal knowledge base; S4: Construct a large-scale model for fine-tuning in the field of power plant safety regulations, and input standardized query data and related knowledge fragments of recall and optimization into the large-scale model for fine-tuning in the field of power plant safety regulations to generate a set of candidate answers that conform to safety regulations; S5: Perform compliance verification and confidence assessment on the candidate answer set, select the answer with the highest confidence as the final response, and feed back the final response and its associated safety data to the user.

2. The intelligent question-and-answer interaction method for power plant safety regulations based on a large model as described in claim 1, characterized in that, In S1, the first preprocessing of the multi-source heterogeneous data of the power station generates a power station safety regulation knowledge graph and a text-speech-image adapted multimodal vector embedding, constructing a power station safety regulation multimodal knowledge base, including: Collect structured power plant safety regulations and equipment operating procedures, unstructured accident cases and maintenance records, as well as multimodal equipment diagrams and operation video frames to obtain multi-source heterogeneous data; Text information from images and video frames in multi-source heterogeneous data is extracted using OCR technology, and then cleaned and normalized to obtain text data. Semantic word segmentation and named entity recognition technologies are used to annotate power plant-specific terms, and entity-relation-attribute triples are established to form a power plant safety regulation knowledge graph. The entity relationship features in text data and power plant safety regulations knowledge graph are mapped to a unified semantic space to generate multimodal vector embeddings; By storing multimodal vector embeddings and power plant safety regulations knowledge graphs in a hybrid storage architecture consisting of a vector database and a graph database, a multimodal knowledge base for power plant safety regulations is obtained.

3. The intelligent question-and-answer interaction method for power plant safety regulations based on a large model as described in claim 1, characterized in that, In S2, receiving the user's input question-and-answer request and performing a second preprocessing on the request to generate standardized query data includes: The system receives user-inputted question-and-answer requests, performs modal parsing, noise filtering, and semantic enhancement on the requests, extracts core entities and scene features, and generates standardized query data. The question-and-answer requests include at least one of text, voice, and image.

4. The intelligent question-and-answer interaction method for power plant safety regulations based on a large model as described in claim 3, characterized in that, The semantic enhancement process specifically includes: For text requests, a dedicated dictionary for the power plant domain is used to optimize word segmentation of the question-and-answer requests after modal parsing and noise filtering; core entities are identified through a pre-trained named entity recognition model, and semantic missing information is supplemented by context to transform fuzzy requests into standardized query data; For voice requests, the data is first converted into text using speech-to-text technology, and then semantic enhancement processing steps are performed for text requests to obtain standardized query data. For image requests, semantic query vectors are generated by combining image features with text descriptions to obtain standardized query data.

5. The intelligent question-and-answer interaction method for power plant safety regulations based on a large model as described in claim 3, characterized in that, In S3, the step of retrieving relevant knowledge fragments from the power plant safety regulations multimodal knowledge base using a hybrid retrieval strategy based on the retrieval enhancement generation mechanism includes: Based on core entities, a preliminary knowledge set that matches the characteristics of the scenario is filtered out through Boolean retrieval, while irrelevant domain data is eliminated; The semantic similarity between the standardized query data and the preliminary knowledge set is calculated by vector semantic retrieval, and the top-k related knowledge fragments are recalled, where k is a positive integer greater than or equal to 2; Based on the entity relationship links of the power plant safety regulations knowledge graph, path scoring and ranking optimization are performed on the top-k knowledge fragments.

6. The intelligent question-and-answer interaction method for power plant safety regulations based on a large model as described in claim 1, characterized in that, In S4, the construction of the fine-tuned large model in the field of power plant safety regulations includes: Construct a dataset for fine-tuning power plant safety regulations; Based on the power plant safety regulation fine-tuning dataset, LoRA lightweight fine-tuning technology is used to freeze the basic parameters of the basic large model and train the parameters of the domain adaptation layer and attention layer of the basic large model. By introducing a safety compliance loss function, the consistency between the generated answer of the large model and the safety regulations is constrained, resulting in a fine-tuned large model for power plant safety regulations. The basic large model is selected from the GPT series, ERNIE or LLaMA series pre-trained models.

7. The intelligent question-and-answer interaction method for power plant safety regulations based on a large model as described in claim 1, characterized in that, In S5, the compliance verification and confidence assessment of the candidate answer set, and the selection of the answer with the highest confidence as the final response, includes: A compliance verification rule base was established based on the power plant safety regulations knowledge graph and the core clauses of the power plant safety regulations. By comparing the candidate answer set with the compliance verification rule base using a semantic matching algorithm, conflict detection and process compliance judgment are performed to obtain a compliant answer set; A pre-set reliability threshold is used to calculate the confidence score based on the semantic fit between the set of compliant answers and the question-and-answer request, the sufficiency of the retrieved evidence, and the matching accuracy of the safety regulations. The score is calculated using a weighted summation method. If the confidence score is greater than or equal to the confidence threshold, the highest confidence score will be selected as the final response. If the confidence score is less than the confidence threshold, the set of compliant answers needs to be supplemented or the user needs to be prompted to refine the question.

8. The intelligent question-and-answer interaction method for power plant safety regulations based on a large model as described in claim 1, characterized in that, The power plant safety regulation intelligent question-and-answer interaction method based on a large model also includes: Incremental updates to the power plant safety regulations multimodal knowledge base are performed, specifically as follows: Real-time collection of revised safety regulations for power plants, operating procedures for newly added equipment, and the latest accident cases yields new data. This data is then updated using a combination of incremental web crawling and manual review to address the multi-source heterogeneous data. The first preprocessing step S1 is performed on the newly added data to generate an incremental knowledge graph and vector embeddings. An incremental update algorithm is used to integrate incremental knowledge graphs and vector embeddings into a multimodal knowledge base for power plant safety regulations, thereby achieving incremental updates of the knowledge base. At the same time, redundancy cleanup and contradiction detection are carried out regularly in the knowledge base, and multi-source data conflicts are resolved based on data source priority rules.

9. The intelligent question-and-answer interaction method for power plant safety regulations based on a large model as described in claim 1, characterized in that, The power plant safety regulation intelligent question-and-answer interaction method based on a large model also includes: Users input follow-up questions and answers, and based on the context of historical questions and answers and the retrieved knowledge, subsequent responses are generated to achieve multi-round continuous interaction; Meanwhile, during the interaction process, user question and answer requests, response content, and operation logs are encrypted and stored, and access permissions are set for hierarchical control.

10. A power plant safety regulation intelligent question-and-answer interactive system based on a large model, characterized in that, The intelligent question-and-answer interaction method for power plant safety regulations based on a large model, as described in any one of claims 1-9, includes: The power plant safety regulations multimodal knowledge base construction module is used to perform the first preprocessing of multi-source heterogeneous data of the power plant, generate a power plant safety regulations knowledge graph and text-speech-image adapted multimodal vector embedding, and construct a power plant safety regulations multimodal knowledge base; The standardized query data generation module is used to receive user-inputted question-and-answer requests, perform a second preprocessing on the requests, and generate standardized query data. The associated knowledge fragment recall and optimization module is used to recall and optimize associated knowledge fragments from the power plant safety regulations multimodal knowledge base based on the retrieval enhancement generation mechanism and through a hybrid retrieval strategy. The candidate answer set generation module is used to build a fine-tuning large model in the field of power plant safety regulations. It inputs standardized query data and related knowledge fragments of recall and optimization into the fine-tuning large model in the field of power plant safety regulations to generate a candidate answer set that conforms to safety regulations. The final response acquisition module is used to perform compliance verification and confidence assessment on the candidate answer set, select the answer with the highest confidence as the final response, and feed the final response and its associated safety data back to the user.

Images