An AI performance-based security management intelligent question and answer system
By building an AI-powered intelligent Q&A system for safety management, the problems of difficult knowledge updates and insufficient intelligence in Q&A in traditional safety management systems have been solved. This has enabled dynamic management and personalized delivery of safety knowledge, improving the intelligence of the Q&A system and the user experience.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- CHINA HYDROELECTRIC ENGINEERING CONSULTING GROUP CHENGDU RESEARCH HYDROELECTRIC INVESTIGATION DESIGN AND INSTITUTE
- Filing Date
- 2026-03-17
- Publication Date
- 2026-06-19
AI Technical Summary
Traditional security management systems cannot achieve real-time dynamic updates and personalized push of knowledge. They have low levels of intelligence in question-and-answer interaction and lack user feedback mechanisms, resulting in outdated knowledge resources, insufficient accuracy of questions and answers, and poor user experience.
A smart question-and-answer system for safety management based on AI efficiency is constructed, including a knowledge resource base layer, an intelligent question-and-answer engine layer, a question-and-answer business control layer, and a multi-terminal interaction adaptation layer. Through multi-channel data collection, dynamic semantic slicing, personalized recommendation algorithms, and a user feedback closed-loop mechanism, dynamic management of the knowledge base and intelligent question-and-answer are achieved.
It enables efficient, accurate, and personalized management of security knowledge, improves the intelligence level and user experience of the question-and-answer system, ensures the timeliness and accuracy of the knowledge base, and continuously optimizes system performance through user feedback.
Smart Images
Figure CN122240775A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of safety management, specifically to a smart question-and-answer system for safety management based on AI performance. Background Technology
[0002] With the increasing complexity and growing demand for intelligent safety production management in enterprises, safety management work faces the challenge of processing massive amounts of data. Traditional safety management information systems mainly rely on pre-set rule bases or simple database queries, where users retrieve relevant documents or standard answers by inputting keywords. However, this model has the following significant drawbacks: 1. Knowledge is isolated and difficult to update: Security-related regulations, standards and internal procedures are frequently updated. Traditional systems struggle to collect and maintain knowledge in real time, resulting in outdated knowledge resources that cannot provide effective support for decision-making.
[0003] 2. Low level of intelligence in question-and-answer interaction: Most existing systems can only handle explicit keyword matching, and cannot understand the user's vague or colloquial question intent. They also have difficulty combining context to conduct multi-turn dialogues, resulting in insufficient accuracy of question and answer and poor user experience.
[0004] 3. Lack of personalization and context awareness: The system cannot provide targeted knowledge push based on the user's role, historical behavior or current business scenario, making it difficult to meet the differentiated knowledge needs of managers at different levels or front-line operators.
[0005] 4. Lack of a closed-loop system operation and optimization mechanism: Traditional systems often focus only on information output, neglecting valuable data such as user feedback and Q&A records. The lack of an evaluation mechanism for the effectiveness of Q&A and the ability to self-optimize based on feedback prevents the system's performance from continuously improving.
[0006] Therefore, how to build an intelligent question-and-answer system that can achieve dynamic knowledge management, possess highly intelligent question-and-answer capabilities, and continuously optimize itself based on user feedback has become a pressing technical challenge in the field of information technology for security management. Summary of the Invention
[0007] In order to achieve dynamic knowledge management, possess highly intelligent question-and-answer capabilities, and continuously optimize based on user feedback, this application provides a security management intelligent question-and-answer system based on AI performance.
[0008] The technical solution adopted by the present invention to solve the above problems is: A security management intelligent question-and-answer system based on AI performance includes: The knowledge resource foundation layer is used for the collection, processing, storage, and dynamic management of security knowledge data; The intelligent question-answering engine layer is used for question-answering processing and personalized knowledge delivery; The question-and-answer business control layer is used for user management, question-and-answer management, knowledge graph operation and maintenance management, and system optimization based on user feedback. A multi-platform interaction adaptation layer is used to provide user access interfaces.
[0009] Furthermore, the knowledge resource foundation layer includes: The safety knowledge collection and processing module is used to collect various safety management-related knowledge resources in the field of engineering construction from multiple channels, and to slice and annotate the collected data using a dynamic semantic slicing and entity annotation fusion algorithm. The intelligent knowledge base module is used for the graph-based construction of knowledge in the security field.
[0010] Furthermore, data collection channels include: internal enterprise database integration, crawling from official policy platforms, industry literature retrieval, and expert experience input.
[0011] Furthermore, the annotation information includes the knowledge point category, difficulty level, applicable job positions, related knowledge points, and policy update time.
[0012] Furthermore, the intelligent question-answering engine layer includes: The distributed model scheduling module is used to manage multiple types of AI models and schedule them based on load balancing. An enhanced question-answering processing module is used to complete question-answering processing based on retrieval enhancement algorithms and natural language processing algorithms; The personalized knowledge push module is used to select suitable knowledge points based on personalized recommendation algorithms and push them to users regularly.
[0013] Furthermore, the AI models include the LLaMA 3×Qwen fusion LLM model, the BGE-M3Embedding model, and the Cross-Encoder Rerank model.
[0014] Furthermore, the retrieval enhancement algorithm adopts enhanced RAG, and the natural language processing algorithm adopts a fusion of PromptEngineering, Few-Shot Learning, and Mind Chain NLP.
[0015] Furthermore, the personalized recommendation algorithm integrates federated learning and deep reinforcement learning, and constructs accurate user profiles based on user positions, knowledge gaps, question-and-answer history, and learning habits to achieve personalized recommendations.
[0016] Furthermore, the question-and-answer business control layer includes: The user identity management module is used for user identity management; The Q&A full-process management module is used for the full-process management of Q&A history, intelligent review, hot topic mining, and data visualization and statistical analysis. The knowledge graph operation and maintenance module is used for the full-process management of the knowledge base; The user feedback handling module is used to collect user satisfaction feedback, opinions and suggestions, and error correction information on the question and answer results. The collected information is classified and analyzed to screen out system optimization points and knowledge supplementation points, and corresponding optimizations are made based on the classification and analysis results.
[0017] Furthermore, the multi-terminal interaction adaptation layer provides a PC-side backend management interface and a mobile APP interface.
[0018] The advantages of this invention compared to existing technologies are as follows: By constructing a knowledge resource foundation layer, it achieves unified collection, standardized processing, and dynamic update management of multi-source heterogeneous security data, ensuring the accuracy, completeness, and timeliness of the knowledge base and providing high-quality data support for upper-layer intelligent applications; by constructing an intelligent question-answering engine layer, it utilizes AI technologies such as natural language processing and deep learning to achieve accurate understanding of users' complex intentions, providing accurate question-answer matching and personalized knowledge push services, thus improving the intelligence level of human-computer interaction; by constructing a question-answering business control layer, it establishes a closed-loop user feedback mechanism, using user behavior, question-answering records, and other data to evaluate system performance, and accordingly optimizes the question-answering model and knowledge graph, achieving continuous learning and evolution of the system. Through the organic integration of multi-layer architecture, a highly efficient, accurate, and evolvable intelligent question-answering system is constructed. Attached Figure Description
[0019] Figure 1 This is an architecture diagram of a security management intelligent question-and-answer system based on AI performance. Detailed Implementation
[0020] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.
[0021] like Figure 1 As shown, a security management intelligent question-and-answer system based on AI performance includes: The knowledge resource foundation layer is used for the collection, processing, storage, and dynamic management of security knowledge data.
[0022] The knowledge resource foundation layer is used for the full-domain collection, intelligent processing, distributed storage and dynamic management of security knowledge data. It provides high-quality and timely data support for the question-and-answer function of the entire system and is the core foundation for the stable operation of the system. It includes a security knowledge collection and processing module and an intelligent knowledge base module.
[0023] The safety knowledge collection and processing module comprehensively collects various safety management-related knowledge resources in the field of engineering construction through multiple channels, including enterprise internal database integration, official policy platform crawling, industry literature retrieval, and expert experience input. It adopts a dynamic semantic slicing and entity annotation fusion algorithm to divide long documents into short segments of 400-800 words each containing complete knowledge points. Combined with knowledge graph association annotation and multi-dimensional intelligent classification technology, it annotates knowledge point categories, difficulty levels, applicable positions, related knowledge points, and policy update times. Compared with traditional slicing and annotation methods, the efficiency is improved by more than 70%, and the annotation accuracy rate reaches more than 99.5%.
[0024] The intelligent knowledge base module adopts a dual-engine collaborative architecture of Elasticsearch and Milvus vector database, combining enhanced RAG retrieval technology and knowledge graph embedding strategy to construct a graph-based structured knowledge base from the preprocessed security knowledge dataset. It is divided into multiple specialized sub-knowledge bases such as technical standards, laws and regulations, emergency response, and job practice, and establishes a four-fold vector index. It supports four methods: semantic search, keyword search, association search, and voice search. The recall and precision rates of knowledge retrieval both reach over 99.2%, and the retrieval response time is shortened to within 40ms. At the same time, it supports administrators to perform full-process visual maintenance operations on the knowledge base, realizing real-time dynamic updates of knowledge and automatic warnings of expired knowledge.
[0025] The intelligent question-answering engine layer is used for question-answering processing and personalized knowledge delivery.
[0026] The intelligent question-answering engine layer integrates multiple high-level fusion AI models, providing the system with core intelligent driving capabilities. It is the core innovative unit for realizing the intelligent question-answering function, including a distributed model scheduling module, an enhanced question-answering processing module, a personalized knowledge push module, and a model adaptive iteration module.
[0027] The distributed model scheduling module adopts a distributed model scheduling cluster architecture, integrating the LLaMA 3×Qwen fusion LLM model, BGE-M3Embedding model, and Cross-Encoder Rerank model. Through API gateway cluster encapsulation and intelligent load balancing scheduling technology, it realizes unified scheduling, real-time monitoring, fault redundancy, and dynamic expansion of various AI models. It supports seamless integration with hundreds of proprietary / open source LLMs and dozens of inference providers. Combined with the LangChain Pro+FastAPI framework, it optimizes the model call chain, improves the question-answering response speed to within 250ms, and achieves model call stability of over 99.98%.
[0028] The enhanced question-answering module is based on enhanced RAG retrieval technology and a natural language processing (NLP) technology that integrates Prompt Engineering, Few-Shot Learning, and CoT (CoT). It can accurately receive user natural language and voice queries, and through multi-step semantic parsing, intent recognition, and ambiguity resolution, generate accurate, concise, and job-relevant answers. It supports multi-turn dialogue memory and contextual association, automatically records question-answering history, and adapts to subsequent follow-up questions. Specifically, it includes: The receiving unit is configured to receive natural language queries or voice queries input by the user; The semantic parsing and intent recognition unit is configured to perform multi-step semantic parsing on the received user query, including but not limited to word segmentation, part-of-speech tagging, dependency parsing and named entity recognition, and perform user intent recognition and ambiguity resolution based on the parsing results; The context management unit is configured to record and maintain context information for multi-turn dialogues, supporting dynamic updates and long-term memory of dialogue history, ensuring that subsequent follow-up questions can be semantically completed and related to understanding based on historical conversations; The retrieval and reasoning units are enhanced and configured to retrieve the most relevant information fragments from the knowledge base using RAG technology, based on user queries and dialogue context. At the same time, by combining Few-Shot learning examples and CoT reasoning logic, the model is guided to gradually generate solutions that conform to the logical chain and business scenario requirements. The answer generation unit is configured to generate accurate, concise, and relevant response texts based on retrieved evidence and reasoning paths, tailored to the specific job or application scenario.
[0029] In addition, it includes a question and answer history management unit, which is configured to automatically record the question and answer history between users and the system, and supports users to perform multi-dimensional operations on the history, including but not limited to viewing, searching by keywords, batch exporting, and bookmarking.
[0030] By deeply integrating enhanced RAG retrieval technology with NLP technology that combines Prompt Engineering, Few-Shot Learning, and CoT (CoT), and combining four retrieval modes, this technology enables natural language interactive question-and-answer, voice question-and-answer, and multi-turn coherent dialogue for safety knowledge. This solves the problems of low efficiency, poor accuracy, and insufficient convenience in existing safety management knowledge retrieval methods. The average user knowledge query and response time is reduced from 20 minutes in the traditional way to within 0.25 minutes, and the retrieval and response accuracy rate reaches over 99.2%, significantly improving the efficiency and convenience of knowledge acquisition for safety management personnel.
[0031] The personalized knowledge recommendation module employs a personalized recommendation algorithm that integrates federated learning, deep reinforcement learning, and user behavior profiling. Without disclosing user privacy data, it constructs precise user profiles based on user job title, knowledge gaps, question-and-answer history, and learning habits, selecting suitable knowledge points and dynamically adjusting the push frequency (1-3 times / week) using an adaptive push mechanism. The pushed content primarily focuses on lightweight exam points, core knowledge points, explanations of common mistakes, and interpretations of the latest policies. Specifically, it includes: The data acquisition unit is configured to acquire user behavior data, including user job title, knowledge gaps, question and answer history, and learning habits. A local model training unit, deployed on the user terminal, is used to perform federated learning tasks. This local model training unit trains a local user interest representation model based on local behavioral data and only uploads encrypted model gradient parameters to the central server. The central server aggregates gradients from all sources to update the global recommendation model, and then distributes the updated global model parameters to each terminal. This mechanism ensures iterative optimization of model performance even when the central server cannot access the original user data. The profile building unit is used to build multi-dimensional user profiles. This unit receives updated global model parameters under the federated learning framework and combines a deep reinforcement learning network to dynamically analyze the user's real-time feedback. The adaptive push unit, based on user profiles, selects the most relevant knowledge points from the knowledge base, including exam points, core knowledge points, explanations of common mistakes, and interpretations of the latest policies.
[0032] Furthermore, the adaptive push unit dynamically adjusts the frequency of knowledge pushes to the same user based on the decision results output by the deep reinforcement learning network. This push frequency is not a fixed value, but is adaptively adjusted according to the user's learning fatigue level, availability of fragmented time, and knowledge acquisition rate reflected in the user profile.
[0033] The question-and-answer business control layer is used for user management, question-and-answer management, knowledge graph operation and maintenance management, and system optimization based on user feedback.
[0034] The question-and-answer business control layer is used to realize the fine-grained processing and control of the entire process of question-and-answer related business logic in the system, supporting the implementation of system functions and efficient operation and maintenance. It includes a user identity control module, a question-and-answer full-process management module, a knowledge graph operation and maintenance module, and a user feedback handling module.
[0035] The user identity management module supports multiple account registration modes, including enterprise account association registration, mobile phone number registration, and third-party authorization registration. It adopts a multimodal adaptive authentication and behavior risk control fusion mechanism that combines biometrics (face / fingerprint) and dynamic tokens. The authentication method is dynamically adjusted based on the account risk level. Four roles are set: system administrator, knowledge administrator, general security personnel, and expert user. Role-based dynamic permission adaptation technology is used to achieve fine-grained dynamic permission allocation. It supports user account security management, personal information maintenance, login log query and abnormal behavior warning. Administrators can realize full lifecycle management of users and data visualization monitoring.
[0036] The Q&A full-process management module enables full-process management of Q&A history, intelligent review, hot topic mining, and data visualization and statistical analysis. It focuses on Q&A pain point analysis, high-frequency question summary, and Q&A quality assessment, automatically filtering out illegal and invalid Q&A content, and generating Q&A performance analysis reports to provide accurate basis for system optimization and knowledge supplementation.
[0037] The knowledge graph operation and maintenance module enables full-process visualized maintenance of specialized sub-knowledge bases. It optimizes the knowledge structure by combining knowledge graph association technology, supports rapid knowledge retrieval, batch import and export, review and release, version management, and knowledge conflict detection. It can automatically identify duplicate and conflicting knowledge in the knowledge base and issue warnings to ensure the professionalism and accuracy of the knowledge base content.
[0038] The feedback management module employs a fusion of sentiment analysis, intent recognition, and keyword extraction technology to automatically collect user feedback on question-and-answer results, suggestions, and error correction information. It performs intelligent classification and analysis, quickly filters system optimization points and knowledge supplementation points, and achieves closed-loop management of feedback-optimization-feedback. It simultaneously provides users with optimization results and supplementary knowledge content, thereby improving the user experience.
[0039] A multi-platform interaction adaptation layer is used to provide user access interfaces.
[0040] The multi-terminal interaction adaptation layer provides a convenient, easy-to-use, and efficient user interface, supports seamless switching between multiple terminals, and meets the question and answer needs of security personnel in different scenarios, including the PC backend management interface and the mobile APP interface.
[0041] The PC-based backend management interface is developed using the advanced architecture of React 18, Next.js, and Ant Design Pro. It adopts a three-section layout of left navigation bar, top toolbar, and right content area, supporting custom layout, theme switching, and large-screen data visualization. It enables full-process system control, real-time data monitoring, and report export. The mobile APP interface is developed using the Flutter 4.0 cross-platform framework and native plugins, adapting to all sizes of mobile devices (phones and tablets). It adopts a bottom menu bar and middle content area layout, integrating gesture operation, AI voice Q&A, offline caching, message push, and other functions. It supports viewing historical Q&A and cached knowledge in offline environments, with an offline usage time of up to 96 hours.
[0042] The mobile app incorporates AI voice Q&A and incremental offline caching to meet the needs of security personnel for Q&A and learning anytime, anywhere, solving the problem of existing knowledge acquisition channels being limited by time and location. At the same time, through reinforcement learning-driven model adaptive optimization and feedback closed-loop management, the system's performance is continuously improved, resulting in a better user experience.
[0043] The system also includes a full-dimensional test optimization module, which adopts a five-dimensional testing system of automated testing, stress testing, anomaly simulation testing, chaos testing, and compatibility testing to conduct full-dimensional and full-scenario testing on the system. Based on the test results, intelligent optimization algorithms are used to precisely optimize each module of the system.
Claims
1. A security management intelligent question-and-answer system based on AI performance, characterized in that, include: The knowledge resource foundation layer is used for the collection, processing, storage, and dynamic management of security knowledge data; The intelligent question-answering engine layer is used for question-answering processing and personalized knowledge delivery; The question-and-answer business control layer is used for user management, question-and-answer management, knowledge graph operation and maintenance management, and system optimization based on user feedback. A multi-platform interaction adaptation layer is used to provide user access interfaces.
2. The AI-based intelligent question-and-answer system for security management according to claim 1, characterized in that, The knowledge resource foundation layer includes: The safety knowledge collection and processing module is used to collect various safety management-related knowledge resources in the field of engineering construction from multiple channels, and to slice and annotate the collected data using a dynamic semantic slicing and entity annotation fusion algorithm. The intelligent knowledge base module is used for the graph-based construction of knowledge in the security field.
3. The AI-based intelligent question-and-answer system for security management according to claim 2, characterized in that, Data collection channels include: internal enterprise database integration, official policy platform crawling, industry literature retrieval, and expert experience input.
4. The AI-based intelligent question-and-answer system for security management according to claim 2, characterized in that, The annotation information includes the knowledge point category, difficulty level, applicable job positions, related knowledge points, and policy update time.
5. The AI-based intelligent question-and-answer system for security management according to claim 1, characterized in that, The intelligent question-answering engine layer includes: The distributed model scheduling module is used to manage multiple types of AI models and schedule them based on load balancing. An enhanced question-answering processing module is used to complete question-answering processing based on retrieval enhancement algorithms and natural language processing algorithms; The personalized knowledge push module is used to select suitable knowledge points based on personalized recommendation algorithms and push them to users regularly.
6. The AI-based intelligent question-and-answer system for security management according to claim 5, characterized in that, The AI models include the LLaMA 3×Qwen fusion LLM model, the BGE-M3 Embedding model, and the Cross-Encoder Rerank model.
7. A security management intelligent question-and-answer system based on AI performance as described in claim 5, characterized in that, The retrieval enhancement algorithm adopts enhanced RAG, and the natural language processing algorithm adopts a fusion of Prompt Engineering, Few-Shot Learning, and Mind Chain NLP.
8. A security management intelligent question-and-answer system based on AI performance according to claim 5, characterized in that, The personalized recommendation algorithm integrates federated learning and deep reinforcement learning, and builds accurate user profiles based on user job, knowledge gaps, question-and-answer history and learning habits to achieve personalized recommendations.
9. A security management intelligent question-and-answer system based on AI efficiency according to claim 1, characterized in that, The Q&A business control layer includes: The user identity management module is used for user identity management; The Q&A full-process management module is used for the full-process management of Q&A history, intelligent review, hot topic mining, and data visualization and statistical analysis. The knowledge graph operation and maintenance module is used for the full-process management of the knowledge base; The user feedback handling module is used to collect user satisfaction feedback, opinions and suggestions, and error correction information on the question and answer results. The collected information is classified and analyzed to screen out system optimization points and knowledge supplementation points, and corresponding optimizations are made based on the classification and analysis results.
10. A security management intelligent question-and-answer system based on AI efficiency according to claim 1, characterized in that, The multi-terminal interaction adaptation layer provides a PC-side backend management interface and a mobile APP interface.