Multi-dimensional risk analysis and early warning method based on intelligent agent

Abstract

The invention discloses a multi-dimensional risk analysis and early warning method based on an intelligent agent. The method comprises the following steps: step 1, constructing a risk knowledge base; individual portraying is carried out on social risk field consultation questions based on ReAct, and personalized answer and derivation strategies are provided for different risk individuals according to the individual portraits; in the process of responding to user consultation, risk individuals involved in the dialogue are automatically identified and deeply analyzed, and a corresponding intelligent analysis report is generated; intelligent early warning agents are synchronously executed in the generation process of the intelligent report, and the intelligent effect of early warning notification is achieved; the risk assessment big language model is optimized, through continuous iterative training, the big language model can better adapt to the requirement of risk assessment, and the prediction accuracy and efficiency of the big language model are improved. According to the invention, high automation of risk analysis is realized.

Description

Technical Field

[0001] The present invention relates to the technical field of big data risk assessment, and in particular to a multi-dimensional risk analysis and early warning method based on an intelligent agent. Background Art

[0002] Traditional monitoring and event analysis systems have problems such as low automation and inaccurate analysis, which limits the accuracy of risk analysis.

[0003] In the existing technology, there are the following major technical problems: 1. Low degree of automation: Traditional technology has a low level of automation in monitoring and event analysis systems, requiring more manual intervention. 2. Decision-making intelligence: The decision-making ability in traditional technology is limited and can only execute preset instructions. 3. Explainability: Risk assessment in traditional technology lacks transparency, making it difficult to understand the reasons behind the decision. 4. Data application: Existing technology may not fully utilize the data accumulation of the big data platform. Summary of the invention

[0004] The purpose of the present invention is to provide a multi-dimensional risk analysis and early warning method based on intelligent agents, which innovates the field of key person monitoring and event analysis by combining large language models (LLM) and intelligent agents (AI-Agent) technology, and realizes a high degree of automation of risk analysis.

[0005] The technical solution adopted by the present invention is: The multi-dimensional risk analysis and early warning method based on intelligent agents includes the following steps: Step 1: Build a risk knowledge base: Use the risk assessment big data platform to collect multi-dimensional data information and extract key features from the multi-dimensional data information to build a social risk knowledge base; Furthermore, the multi-dimensional data information in step 1 includes key personnel, historical event data, petition information, public opinion, and 12345 complaint information.

[0006] Furthermore, in step 1, the autoregressive integrated moving average model (ARIMA) and TF-IDF (TermFrequency-Inverse Document Frequency) are used to extract customized key features from the data, thereby realizing a knowledge base specifically built for social risks.

[0007] Specifically, ARIMA is combined with deep learning models (such as long short-term memory networks (LSTM) or gated recurrent units (GRU). ARIMA is first used to perform preliminary data processing and feature extraction, and then the extracted features are input into the deep learning model for more complex feature and association extraction, summarizing the attributes of the features. At the same time, external variables related to time series are introduced, such as basic personal information, family situation indicators, key personnel indicators, etc. By using these external variables as additional inputs, the model's interpretation of the data is enhanced.

[0008] Step 2, knowledge-based intelligent question answering: Based on ReAct, the individuals or events mentioned in the social risk field questions are deeply analyzed to conduct individual profiling, and based on the individual profiling, personalized answers and derivation strategies are provided for different risk individuals; Specifically, the ReAct framework is customized and improved according to the special and historical data of social risks and specific user groups, so as to provide more in-depth and professional knowledge query and derivation strategies for frequently consulted questions in the field of social risks. Specifically, the semantic big model (ReAct framework) is fine-tuned using text data in the field of social risks to make the model familiar with the terminology, concepts and logic in this field. At the same time, the model's hyperparameters, learning rate, and number of iterations are adjusted multiple times to obtain better performance.

[0009] Furthermore, the individual portrait includes background, involved events, knowledge level, and marital status.

[0010] Step 3, Report Agent: In the process of responding to user inquiries, the agent not only provides instant answer services, but also automatically identifies and deeply analyzes the risk individuals involved in the conversation, and seamlessly activates the report agent to customize intelligent analysis reports for the corresponding risk individuals; Specifically, the prompt is designed in a targeted manner: The present invention develops more refined and targeted prompts to guide large models to generate analysis reports with more insight and unique perspectives. Customize exclusive question frameworks and prompt information according to different types of risk individuals. Continuously test and optimize the expression and logical structure of prompts to stimulate the innovation ability of large models. Conduct A / B tests to compare the quality and value of reports generated under different prompts.

[0011] In addition, the system is also compatible with the functions of the risk assessment platform, can receive early warning information, and quickly generate corresponding intelligent reports for decision-making reference.

[0012] Step 4, early warning agent: The intelligent early warning agent is executed synchronously in the process of generating intelligent reports. The intelligent early warning agent realizes the intelligent effect of early warning notification by combining large models with small models.

[0013] First, based on a customized set of instructions (prompt), the large model is queried to obtain the risk level and warning information, and then the warning agent is called to use a multi-dimensional warning notification algorithm to implement the calling of different interfaces.

[0014] Furthermore, the dimensional warning notification algorithm in step 4 specifically includes the following steps: Step 4-1, automatically classify risk data into categories through intelligent risk assessment algorithms, mainly into: safety risk, public security risk, event risk, regional event risk, personnel risk, etc., and add category labels to the risk data; Step 4-2, by default, the basic algorithm model is used to analyze risk data of different dimensions (current time, individual location, historical behavior, etc.), and the analysis results (risk level, category label, notification urgency) form a multi-dimensional risk warning notification algorithm model; Step 4-3: Based on the basic algorithm model, comprehensively analyze the thresholds of different dimensions of risk warning conditions under each category, and assign category labels to the generated new algorithm model according to the category information; Step 4-4, through the data management platform, uniformly arrange for the new algorithm models under different categories to be updated to the corresponding early warning notification operator library; Step 4-5, repeat step 4-1 to step 4-4 until the optimal matching of risk warning & category & data is completed, so as to classify the newly generated risk notification data and adapt the same label.

[0015] Specifically, the adaptability to specific categories will be stronger. Through the above embodiments, the optimal matching of risk warning, category and data is achieved. The newly generated risk notification data is also divided into categories, and algorithms with the same labels are adapted to ensure the actual application effect of the algorithm.

[0016] Steps 4-6: When there is new warning notification data, the new data points are automatically categorized and labeled, and the algorithm matching the new data is directly called for analysis, greatly improving the accuracy of the algorithm.

[0017] Step 5: Feedback, learning, and optimization of the large language model for risk assessment: The performance data of the large language model when processing risk assessment tasks will be collected and analyzed. Based on the analysis results, the parameters of the large language model will be adjusted and new features or variables will be introduced to enrich the model's input information and improve its predictive capabilities. Through continuous iterative training, the large language model will better adapt to the needs of risk assessment and improve its predictive accuracy and efficiency.

[0018] Furthermore, the performance data in step 5 includes feedback in multiple dimensions such as accuracy, response time, and warning effect. According to the analysis results, the parameters of the large language model are adjusted. For example, if it is found that the model performs poorly in certain types of risk assessment tasks, the amount of training of relevant data can be increased in a targeted manner. Introduce new features or variables to enrich the input information of the model and improve the prediction ability. This feedback information will be used to guide further learning and optimization of the model. Through continuous iterative training, the large language model will better adapt to the needs of risk assessment and improve its prediction accuracy and efficiency.

[0019] The present invention adopts the above technical solution, and achieves high automation by integrating large language models (LLM) and intelligent agents (AI-Agent) technology, reduces dependence on manual work, and improves efficiency and timeliness. The AI-Agent in the system can not only generate reports, but also make automated decisions based on the results of model analysis, which improves the accuracy and speed of decision-making. The system provides the reasoning process of decision-making through the ReAct framework, enhances the interpretability of the system, and facilitates auditing and trust establishment. By deeply integrating the rich information of the risk assessment big data platform, the present invention implements professional fine-tuning of the basic semantic big model, and especially optimizes the ability of multi-dimensional risk analysis and the alignment of the corresponding Agent. By combining large language models (LLM) and intelligent agents (AI-Agent) technology, the present invention innovates the field of key person monitoring and event analysis and achieves high automation of risk analysis. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] The present invention is further described in detail below with reference to the accompanying drawings and specific embodiments; Figure 1 It is a flow chart of the multi-dimensional risk analysis and early warning method based on intelligent agent of the present invention; Figure 2 This is a schematic diagram of the ReAct framework of the present invention; Figure 3 A schematic diagram of the generation principle of the intelligent analysis report of the present invention; Figure 4 Schematic diagram of the process flow of intelligent early warning notification of the present invention; Figure 5 Schematic diagram of the architecture of the multi-dimensional risk warning notification algorithm of the present invention. DETAILED DESCRIPTION

[0021] In order to make the purpose, technical solutions and advantages of the embodiments of the present application clearer, the technical solutions in the embodiments of the present application will be clearly and completely described below in conjunction with the drawings in the embodiments of the present application.

[0022] In order to innovate the technology and practice in the field of risk analysis, the advanced large-scale language model and AI-Agent technology are combined to achieve high automation and intelligence of risk monitoring. Large language models and smaller-scale machine learning models are used at the same time, and their advantages and characteristics are integrated and complemented. This system aims to improve the efficiency, accuracy and transparency of risk analysis, and has the ability of dynamic adaptability and continuous learning to respond to and effectively manage complex and changing risk environments in real time, ultimately improving social security, maintaining stability, and promoting safe development. This will greatly improve the efficiency and effectiveness of governments and security management departments in preventing and responding to high-risk events.

[0023] like Figures 1 to 5 As shown in Figure 1, the present invention discloses a multi-dimensional risk analysis and early warning method based on intelligent agents. With the help of text data in the field of social risks, the semantic big model is fine-tuned in a targeted manner, so that the model is familiar with the terminology, concepts and logic in this field, thereby constructing a big language model dedicated to multi-dimensional risk assessment and early warning services. The big language model conducts a multi-dimensional assessment of a specific high-risk individual and intelligently generates intelligent BI reports and early warnings; specifically, the following steps are included: Step 1, build a risk knowledge base: use the risk assessment big data platform to collect information on key personnel, historical event data, petition information, public opinion, 12345 complaints, etc., and use the autoregressive integrated moving average model (ARIMA) and TF-IDF (TermFrequency-Inverse Document Frequency) to extract customized key features from the data, thereby realizing a knowledge base specifically built for social risks.

[0024] Different from traditional ARIMA, ARIMA is combined with deep learning models (such as long short-term memory network LSTM or gated recurrent unit GRU). ARIMA is first used to perform preliminary data processing and feature extraction, and then the extracted features are input into the deep learning model for more complex feature and association extraction, summarizing the attributes of the features. At the same time, external variables related to time series are introduced, such as basic personal information, family situation indicators, key personnel indicators, etc. By using these external variables as additional inputs, the model's interpretation of the data is enhanced.

[0025] In the calculation of word frequency using TF-IDF (Term Frequency-Inverse Document Frequency), the present invention customizes TF-IDF weights based on special and historical data of social risks and specific user groups. For example, words such as threat, revenge on society, and fighting have higher weights in events, etc. This allows different importance assessments to be made for the same words.

[0026] Step 2, knowledge-based agent question answering: Customize and improve the ReAct framework based on the special and historical data of social risks and specific user groups. Provide more in-depth and professional knowledge query and derivation strategies for frequently consulted questions in the field of social risks.

[0027] In the implementation of ReAct: We used text data from the social risk field to fine-tune the semantic model in a targeted manner, so that the model is familiar with the terminology, concepts, and logic of the field. At the same time, we adjusted the model's hyperparameters, learning rate, number of iterations, etc. multiple times to achieve better performance.

[0028] Realize the integration of knowledge base: Based on LangChain, the large model can interact with the risk knowledge base and use the relationships and information in the graph to provide more accurate and in-depth answers.

[0029] User portrait refinement: The big model of individuals or events mentioned in the consultation on social risk issues will be analyzed more deeply based on ReAct to understand their background, involved events, knowledge level, marital status, etc. Based on the individual portrait, personalized answers and derivation strategies are provided for individuals with different risks.

[0030] Step 3, Report Agent: Figure 3 As shown in the figure, in the process of responding to user inquiries, the intelligent agent not only provides instant answer services, but also automatically identifies and deeply analyzes the risk individuals involved in the conversation. Then, the system will seamlessly activate the report intelligent agent to customize a detailed intelligent analysis report for these risk individuals.

[0031] Targeted design of prompts: This invention develops more refined and targeted prompts to guide large models to generate more insightful and unique analysis reports. Customize exclusive question frameworks and prompt information according to different types of risk individuals. Continuously test and optimize the expression and logical structure of prompts to stimulate the innovation ability of large models. Conduct A / B testing to compare the quality and value of reports generated under different prompts.

[0032] In addition, the system is also compatible with the functions of the risk assessment platform, can receive early warning information, and quickly generate corresponding intelligent reports for decision-making reference.

[0033] Step 4, warning agent: Figure 4 As shown in Figure 5, in the process of generating intelligent reports, the intelligent early warning agent will be executed synchronously. The intelligent agent realizes the intelligent effect of early warning notification by combining large models with small models.

[0034] First, based on a customized set of instructions (prompt), the large model is queried to obtain the risk level and warning information, and then the warning agent is called to use a multi-dimensional warning notification algorithm to implement the calling of different interfaces.

[0035] like Figure 5 As shown, the multi-dimensional early warning notification algorithm multi-dimensional training and management method specifically includes: (1) Automatically classify risk data into categories through intelligent risk assessment algorithms, mainly into: safety risk, public security risk, event risk, regional event risk, personnel risk, etc., and add category labels to the risk data; (2) By default, the basic algorithm model is used to analyze risk data of different dimensions (current time, individual location, historical behavior, etc.), and the analysis results (risk level, category label, notification urgency) form a multi-dimensional risk warning notification algorithm model; (3) Based on the basic algorithm model, comprehensively analyze the thresholds of different dimensions of risk warning conditions under each category, and assign category labels to the generated new algorithm model based on category information; (4) The data management platform is used to uniformly arrange for the update of new algorithm models under different categories to the corresponding early warning notification operator library; (5) By looping the above process, the algorithm model will have a stronger ability to adapt to specific categories. The above embodiments can achieve the optimal matching of risk warning, category, and data. The newly generated risk notification data can also be classified and the algorithm with the same label can be adapted to ensure the actual application effect of the algorithm.

[0036] (6) When there is new warning notification data, the data point is automatically categorized and labeled, and the matching algorithm is directly called for analysis, greatly improving the accuracy of the algorithm.

[0037] Step 5: Feedback, learning, and optimization of the risk assessment large language model: The system will collect and analyze the performance data of the large language model when processing risk assessment tasks, including feedback from multiple dimensions such as accuracy, response time, and warning effect. According to the analysis results, the parameters of the large language model are adjusted. For example, if it is found that the model performs poorly in certain types of risk assessment tasks, the amount of training on the relevant data can be increased in a targeted manner. New features or variables are introduced to enrich the input information of the model and improve the prediction ability. This feedback information will be used to guide further learning and optimization of the model. Through continuous iterative training, the large language model will better adapt to the needs of risk assessment and improve the accuracy and efficiency of its predictions.

[0038] Based on the large language model, multi-dimensional risk warnings such as key personnel and events are fine-tuned, and individual behavior patterns, social networks, historical records, etc. are analyzed. Agents are used to align the large model with intelligent reports and intelligent warnings. The present invention improves the accuracy and comprehensiveness of multi-dimensional risk assessments. The semantic large model is given stronger cognitive and execution capabilities, enabling it to demonstrate human-like intelligent behaviors in complex tasks, rather than just executing preset instructions.

[0039] The present invention constructs a personalized risk assessment knowledge base for each individual being assessed, and relies on a large language model to generate a risk score that reflects the potential threat level to society. At the same time, through the thinking track of ReAct, the system of the present invention can provide the reasoning process behind the decision, increase the transparency of the system, and facilitate auditing and trust establishment. In the intelligent early warning notification module, both large-scale models and small-scale models are used, and their advantages and characteristics are integrated and complementary. The present invention relies on the individual knowledge base. According to the unique background and behavior pattern of each individual, the system can generate more accurate and detailed risk scores, avoiding the misjudgment or omission that may be caused by the "one-size-fits-all" assessment method. The thinking track provided by the ReAct framework makes the decision-making process clear and visible, which not only enhances the credibility of the system itself, but also facilitates the review and verification of regulators and stakeholders to ensure the fairness and rationality of the decision. The more focused and detailed characteristics of the small model are used in the early warning notification agent, and specific and detailed problem handling can be carried out for the early warning method.

[0040] The present invention adopts the above technical solution. By integrating the large language model (LLM) and AI-Agent, the system achieves high automation, reduces the need for manual intervention, greatly improves the efficiency and timeliness of risk monitoring and event analysis, and ensures that relevant departments can respond quickly to various potential risks. The use of AI-Agent can not only generate detailed reports, but also automatically make decisions based on the analysis results of the model, which improves the accuracy and speed of decision-making, especially when dealing with complex events, it can provide more accurate risk assessment and early warning. Using the ReAct framework, the system can provide the reasoning process behind the risk assessment, enhance the interpretability of the decision, increase the transparency of the system, enable users to understand the basis of the decision, and facilitate review and trust. The application in social security monitoring and event early warning helps to enhance the sense of social security, reduce the occurrence of malicious events, and maintain social stability and security. This plays an important role in preventing crime, terrorist activities, public riots, etc. Based on the combination of large and small models of the ReAct architecture, the adaptability to new information and the ability to handle complex events are enhanced. In addition, through the closed-loop feedback system, the system can continuously optimize the model performance and decision quality, and continuously improve its own effectiveness over time.

[0041] Obviously, the described embodiments are part of the embodiments of the present application, rather than all of the embodiments. In the absence of conflict, the embodiments in the present application and the features in the embodiments can be combined with each other. The components of the embodiments of the present application generally described and shown in the drawings here can be arranged and designed in various different configurations. Therefore, the detailed description of the embodiments of the present application is not intended to limit the scope of the application claimed for protection, but merely represents the selected embodiments of the present application. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians of the art without making creative work are within the scope of protection of the present application.

Claims

1. A multi-dimensional risk analysis and early warning method based on an intelligent agent, characterized by: It includes the following steps: Step 1: Build a risk knowledge base: Use the risk assessment big data platform to collect multi-dimensional data information and extract key features from the multi-dimensional data information to build a social risk knowledge base; Step 2, knowledge-based intelligent question answering: Based on ReAct, the individuals or events mentioned in the social risk field questions are deeply analyzed to conduct individual profiling, and based on the individual profiling, personalized answers and derivation strategies are provided for different risk individuals; Step 3, Report Agent: In the process of responding to user inquiries, the agent not only provides instant answer services, but also automatically identifies and deeply analyzes the risk individuals involved in the conversation, and seamlessly activates the report agent to customize intelligent analysis reports for the corresponding risk individuals; Step 4, early warning agent: The intelligent early warning agent is executed synchronously in the process of generating intelligent reports. The intelligent early warning agent achieves the intelligent effect of early warning notification by combining large models with small models: Step 5, optimization of the large language model for risk assessment: collect and analyze the performance data of the large language model when processing risk assessment tasks; adjust the parameters of the large language model based on the analysis results, and introduce new features or variables to enrich the model's input information to improve prediction capabilities; through continuous iterative training, the large language model will better adapt to the needs of risk assessment and improve its prediction accuracy and efficiency.

2. The agent-based multi-dimensional risk analysis and early warning method according to claim 1, characterized in that: The multi-dimensional data information in step 1 includes key personnel, historical event data, petition information, public opinion, and 12345 complaint information.

3. The agent-based multi-dimensional risk analysis and early warning method according to claim 1, characterized in that: In step 1, the autoregressive integrated moving average model ARIMA and TF-IDF are used to extract customized key features from the data, thereby realizing a knowledge base specifically built for social risks.

4. The agent-based multi-dimensional risk analysis and early warning method according to claim 3, characterized in that: Combine ARIMA with deep learning models; first use ARIMA to perform preliminary data processing and feature extraction, and then input the extracted features into the deep learning model for more complex feature and association extraction in order to summarize the attributes of the features; at the same time, introduce external variables related to time series to enhance the model's interpretation of the data. External variables include basic personal information, family situation indicators, and key personnel indicators.

5. The agent-based multi-dimensional risk analysis and early warning method according to claim 1, characterized in that: The individual portrait in step 2 includes background, involved events, knowledge level, and marital status.

6. The agent-based multi-dimensional risk analysis and early warning method according to claim 1, characterized in that: The warning notification in step 4 specifically includes the following steps: Step 4-1, automatically classify risk data into categories through intelligent risk assessment algorithms: safety risk, public security risk, event risk, regional event risk, and personnel risk, and add category labels to the risk data; Step 4-2, by default, uses the basic algorithm model to analyze risk data of different dimensions, and the analysis results form a multi-dimensional risk warning notification algorithm model; Step 4-3: Based on the basic algorithm model, comprehensively analyze the thresholds of different dimensions of risk warning conditions under each category, and assign category labels to the generated new algorithm model according to the category information; Step 4-4, through the data management platform, uniformly arrange for the new algorithm models under different categories to be updated to the corresponding early warning notification operator library; Step 4-5, repeat steps 4-1 to 4-4 until the optimal matching of risk warning, category and data is completed; Steps 4-6: When there is new warning notification data, the new data points are automatically categorized and labeled, and the algorithm matching the new data is directly called for analysis, greatly improving the accuracy of the algorithm.

7. The agent-based multi-dimensional risk analysis and early warning method according to claim 1, characterized in that: The performance data in step 5 includes feedback in multiple dimensions such as accuracy, response time, and warning effect.

Images