Ontology-based and multi-agent risk early warning method and system

Abstract

The application relates to a risk early warning method and system based on an ontology and a multi-agent system. A risk knowledge graph based on an ontology model is constructed or acquired; risk data is collected by a plurality of agents connected with the risk knowledge graph, abnormality detection and risk index analysis, risk identification and attribution analysis, and risk assessment are performed, and finally, early warning processing is performed according to the risk assessment result. The application deeply integrates the ontology and the multi-agent technology, forms an intelligent risk early warning method and system which has a unified knowledge base and distributed processing capability, and is especially suitable for risk monitoring, analysis, assessment and early warning of a complex system.

Description

Technical Field

[0001] This invention relates to the fields of risk management, knowledge engineering, and artificial intelligence, specifically to a risk warning method and system based on ontology and multi-agent system (MAS), which is particularly suitable for risk monitoring, analysis, assessment, and early warning of complex systems. Technical Background The risks faced by modern systems are characterized by multiple sources, interconnectedness, and dynamic evolution.

[0002] Existing risk warning methods include those based on statistical models. These methods collect historical risk data and establish statistical models to predict the probability of future risks. However, they are limited by their reliance on large amounts of historical data and their insufficient ability to predict new and rare risks.

[0003] Existing technologies also include risk warning methods based on expert systems. These methods transform the knowledge and experience of domain experts into a rule base, using rule-based reasoning for risk identification and warning. However, these rules are derived from the experience of domain experts, inevitably containing subjectivity, and are difficult and time-consuming to obtain, exhibiting a time lag.

[0004] Existing technologies include ontology-based financial risk knowledge representation methods and knowledge graph-based risk management methods. However, the former mainly focuses on the static representation of knowledge and lacks dynamic risk monitoring and early warning mechanisms; the latter is mainly used for risk correlation analysis and visualization, and similarly lacks proactive risk monitoring and early warning capabilities.

[0005] Ontology, as a knowledge representation and modeling method, can explicitly define domain concepts, attributes, and their relationships, providing a foundation for knowledge sharing and semantic interoperability among heterogeneous systems. Ontology can construct structured knowledge graphs, enabling the formal representation and reasoning of risk knowledge. However, purely ontological methods lack initiative and adaptability, making it difficult to cope with dynamic risk environments.

[0006] Multi-Agent Systems (MAS) consist of multiple autonomous agents, each capable of sensing the environment, making autonomous decisions, and executing actions to complete complex tasks through communication and collaboration. While MAS technology offers advantages in distributed problem-solving and parallel processing, existing MAS risk warning systems lack a unified knowledge representation framework, resulting in low efficiency in inter-agent communication and collaboration.

[0007] Currently, there is no systematic approach that deeply integrates ontology and multi-agent technology for risk warning. Summary of the Invention

[0008] The purpose of this invention is to provide a risk warning method based on ontology and multi-agent systems to solve the above-mentioned problems in the prior art.

[0009] The risk warning method based on ontology and multi-agent systems of the present invention includes the following steps: Construct or acquire a risk knowledge graph based on an ontology model; Risk data is collected by an intelligent agent connected to a risk knowledge graph, and semantic annotation is performed based on the ontology model. Anomaly detection is performed on the labeled risk data by another intelligent agent connected to the risk knowledge graph, and risk indicator analysis is performed based on the risk knowledge graph. When an anomaly or risk indicator is detected that exceeds a set threshold, another intelligent agent connected to the risk knowledge graph performs risk identification and attribution analysis based on the risk knowledge graph. Risk assessment is conducted by another agent connected to the risk knowledge graph, based on the risk identification and attribution analysis results and according to the risk knowledge graph's evaluation criteria. Another agent connected to the risk knowledge graph performs early warning processing based on the risk assessment results and the early warning rules of the risk knowledge graph.

[0010] In a practical application, the present invention optimizes the early warning method and updates the risk knowledge graph by learning feedback information through another learning optimization agent connected to the risk knowledge graph.

[0011] In a practical application, this invention identifies risk concepts by analyzing risk domain characteristics and needs, defines concept attribute features and semantic relationships between concepts to construct the ontology model; based on the ontology model, multi-source heterogeneous risk information, including historical risk data, expert knowledge, and business rules, is transformed into entities and relationships in a knowledge graph; risk knowledge is extracted through entity recognition, relationship extraction, and attribute filling methods, and organized and stored according to the ontology model specification to form a risk knowledge graph.

[0012] In a practical application, this invention also analyzes the relationships in the risk knowledge graph during risk identification, assesses the transmission path and scope of the risk, and identifies other entities that may be affected.

[0013] In a practical application, the early warning processing described in this invention also includes obtaining risk response measures from a risk knowledge graph.

[0014] In a practical application, the risk assessment described in this invention includes a risk assessment at the current moment, as well as a prediction of the future evolution trend of the risk.

[0015] In a practical application, the intelligent agents of this invention communicate with each other through message passing, and the message content is represented by the concepts and relationships defined in the risk knowledge graph.

[0016] This invention also provides a risk warning system based on ontology and multi-agent systems, including... A risk knowledge graph constructed based on an ontology model; A multi-agent system connected to the risk knowledge graph includes A data acquisition agent collects risk data and performs semantic annotation based on the ontology model. The data analysis agent performs anomaly detection on the risk data labeled by the data collection agent and performs risk indicator analysis based on the risk knowledge graph definition. The risk identification and attribution analysis agent performs risk identification and attribution analysis based on the risk knowledge graph when the data analysis agent detects anomalies or risk indicators exceed a set threshold. The risk assessment agent performs risk assessments based on the risk identification and attribution analysis results of the risk identification and attribution analysis agent, according to the assessment criteria of the risk knowledge graph. The early warning decision-making agent performs early warning processing according to the early warning rules of the risk knowledge graph based on the risk assessment results of the risk assessment agent. The learning agent optimizes itself by learning from feedback information to improve early warning methods and update the risk knowledge graph.

[0017] The present invention also provides an electronic device, including a processor and a memory storing computer program instructions; the processor executes the computer program instructions to implement the steps in the method of the present invention described above.

[0018] The present invention also provides a computer-readable storage medium storing computer program instructions, which, when executed by a processor, implement the steps of the method of the present invention described above.

[0019] This invention deeply integrates ontology with multi-agent technology to form an intelligent risk warning method and system that has both a unified knowledge base and distributed processing capabilities.

[0020] The present invention has significant beneficial effects due to the use of the following technical means: 1. Ontology-based knowledge graph construction method This invention does not simply use ontology as a data model, but rather applies ontology as a methodology for knowledge engineering, systematically guiding the design and construction of knowledge graphs. By clearly defining the conceptual framework, semantic relationships, and reasoning rules of the risk domain through ontology, it ensures that the knowledge graph possesses clear semantics and rigorous logic. The knowledge graph not only stores data, but more importantly, it stores knowledge, including conceptual knowledge, relational knowledge, rule knowledge, and case knowledge. This ontology-based knowledge graph provides a unified semantic foundation for multi-agent systems, enabling different agents to understand and process information within the same semantic framework.

[0021] 2. Functional division of labor and cooperation mechanism among multiple agents Based on the characteristics of risk early warning tasks, this invention designs intelligent agents with multiple specialized functions, including risk data collection, data analysis, risk identification and attribution analysis, risk assessment, early warning decision-making, and learning optimization. Each intelligent agent has clearly defined responsibilities and professional capabilities, enabling it to autonomously complete its assigned sub-tasks. Intelligent agents communicate with each other through ontology-based message passing, with message content represented using concepts and relationships defined in a knowledge graph to ensure communication accuracy. Collaboration between intelligent agents is not simply data transfer, but rather knowledge sharing and collaborative reasoning; multiple agents can jointly analyze complex problems, forming a more powerful problem-solving capability than a single intelligent agent.

[0022] 3. Attribution analysis: the causal reasoning ability of the agent. Attribution analysis is a crucial component of risk management, but traditional methods often fall short in deeply analyzing the causes of risks. This invention provides a knowledge foundation for attribution analysis by explicitly representing causal relationships in a knowledge graph. The attribution analysis agent can reason within a causal network, identifying not only direct causes but also tracing root causes, and analyzing not only single factors but also the interactions of multiple factors. The results of attribution analysis are not only used to understand current risks but also fed back into the knowledge graph, enriching causal knowledge and creating a cycle of continuous improvement.

[0023] 4. Multi-dimensional risk assessment methods The risk assessment method of this invention is not a single-dimensional score, but a comprehensive evaluation from multiple dimensions such as probability, impact, and urgency. The assessment criteria and weight configurations are stored in a knowledge graph and can be flexibly configured according to different risk types and application scenarios. The risk assessment agent not only assesses the current state but also predicts future trends, providing forward-looking information for early warning decisions. This multi-dimensional assessment method makes risk assessment more comprehensive and accurate, avoiding the one-sidedness of single-indicator assessments.

[0024] 5. Learn and optimize the knowledge evolution mechanism of intelligent agents. This invention is not a static knowledge system, but a dynamic system capable of continuous learning and evolution. The learning and optimization agent learns from feedback during system operation and from new risk cases, continuously optimizing system parameters and updating the knowledge graph. The evolution of the knowledge graph follows the specifications of the ontology model, ensuring the consistency of the evolved knowledge graph. This learning and evolution mechanism enables the system to adapt to constantly changing risk environments, and the accuracy and completeness of knowledge continuously improve over time.

[0025] 6. Tight integration of intelligent agents and knowledge graphs In this invention, the intelligent agent is not an external program independent of the knowledge graph, but rather an intelligent component deeply integrated with it. Each intelligent agent can directly access and manipulate the knowledge graph, utilizing the query capabilities of the graph database and the reasoning capabilities of the ontology reasoning engine. The agent's decision-making process is based on the reasoning process of the knowledge graph, and the decision results are updated in the knowledge graph. This tight integration provides the intelligent agent with powerful knowledge support, while also making the knowledge graph a living knowledge base, constantly being used and updated.

[0026] 7. Explainable early warning mechanism The early warning process based on knowledge graphs and ontology reasoning is transparent and traceable. Early warning information not only informs users of the existence and severity of risks but also provides detailed attribution analysis, impact assessments, and response recommendations. Users can query the reasoning path of the early warning to understand how the system identifies, assesses, and makes decisions about risks. This interpretability is crucial for risk management decision-making, enabling users to understand and trust the system's early warning results.

[0027] 8. This invention has good scalability and versatility.

[0028] The modular structure of knowledge graphs facilitates the expansion of knowledge into new risk domains. The multi-agent architecture facilitates the addition of new agents or the expansion of the functionality of existing agents. The methodological framework of this invention can be applied to risk warning in multiple fields such as finance, industry, supply chain, and cybersecurity. Attached Figure Description

[0029] Figure 1 This is a roadmap for risk warning technology based on ontology and multi-agent systems as described in this invention.

[0030] Figure 2 This is a schematic diagram of the process for constructing a risk knowledge graph based on ontology as described in this invention.

[0031] Figure 3 This is a schematic diagram of the structure of the risk knowledge graph described in this invention.

[0032] Figure 4 This is a schematic diagram of the multi-agent system architecture described in this invention.

[0033] Figure 5 This is a schematic diagram of the inter-agent communication and cooperation mechanism described in this invention.

[0034] Figure 6 This is a flowchart of the risk monitoring and data processing for intelligent agent collaboration described in this invention. Figure 7 This is a flowchart of the knowledge graph-based risk identification and attribution analysis described in this invention. Figure 8 This is a flowchart of the multi-dimensional risk assessment process described in this invention. Figure 9 The flowchart of the intelligent early warning decision-making and response described in this invention. Figure 10 This is a flowchart of the knowledge learning and graph evolution process described in this invention. Specific implementation method The present invention will be further described below with reference to the accompanying drawings and specific embodiments, but this does not constitute any limitation on the present invention.

[0035] The method flow of this invention is as follows: Figure 1 As shown, it includes the following steps: Step S1: Construct a risk knowledge graph based on ontology This invention constructs an ontology model for the risk domain using an ontology-based approach, and builds a risk knowledge graph based on this model.

[0036] like Figure 2 As shown, this invention identifies core risk concepts by analyzing the characteristics and needs of the risk domain, including risk events, risk factors, risk indicators, risk levels, risk consequences, risk subjects, risk objects, and risk transmission paths. The identified concepts are then organized hierarchically, establishing a classification system and inheritance relationships to form a hierarchical structure from abstract to concrete. For example, risk events are divided into primary categories such as credit risk, market risk, and operational risk, and credit risk is further subdivided into secondary categories such as default risk and downgrade risk.

[0037] Based on the conceptual framework, semantic relationships between concepts are defined. Causal relationships represent the causal link between risk factors and risk events, such as insufficient liquidity leading to solvency risk. Correlation relationships represent the mutual influence between different risk events, such as the interaction between market risk and credit risk. Compositional relationships represent the inclusion relationship between composite risks and sub-risks. Temporal relationships represent the time sequence and evolution path of risk events. Degree relationships quantify the intensity of the impact of risk factors on risk events. Furthermore, the relationships between risk subjects and risk events, and the mapping relationship between risk indicators and risk assessments are also defined.

[0038] Each risk concept is defined with its attribute characteristics, including attribute name, data type, value range, default value, and constraints. For example, for the risk indicator concept, attributes such as indicator name, indicator value, unit, threshold, and weight are defined. Attribute constraints ensure the consistency and integrity of data in the knowledge graph. The ontology model is formally represented using the ontology description language OWL, and entities, attributes, and relations in the knowledge graph are stored in the form of RDF triples, forming a semantic network that can be understood and processed by machines.

[0039] This invention, based on an ontology model, transforms multi-source heterogeneous risk information, such as historical risk data, expert knowledge, and business rules, into entities and relationships within a knowledge graph. Through techniques like entity recognition, relation extraction, and attribute filling, risk knowledge is extracted from unstructured and semi-structured data and organized and stored according to ontology model specifications. The final result is a risk knowledge graph containing rich risk knowledge and clear semantics, providing a knowledge foundation for subsequent multi-agent systems, such as… Figure 3 As shown.

[0040] Step S2: Construct a multi-agent system architecture Based on the risk knowledge graph constructed in step S1, a multi-agent system architecture is designed and built. For example... Figure 4 As shown, the architecture includes multiple agents with different functions: a data acquisition agent that collects risk-related data from various data sources in real time; a data analysis agent that performs in-depth analysis of the collected data; a risk identification and attribution analysis agent specifically responsible for risk identification and attribution analysis; a risk assessment agent responsible for comprehensively evaluating the severity and scope of risk; an early warning decision agent that makes early warning decisions based on the risk assessment results; and a learning and optimization agent responsible for the continuous learning and optimization of the system. Each agent can access and utilize a unified risk knowledge graph to complete risk early warning tasks through collaborative work. Figure 5 As shown, an agent communication mechanism is established in a multi-agent system, where agents exchange information and collaborate through message passing. Communication messages employ ontology-based semantic representation to ensure that different agents can accurately understand the message content. An agent coordination mechanism is established, including task allocation, conflict resolution, and resource scheduling, to ensure orderly collaboration among multiple agents and avoid duplication of work and resource waste.

[0041] Step S3: Risk monitoring and data processing for agent collaboration In step S3, the multi-agent system of the present invention begins to work collaboratively to perform real-time risk monitoring and data processing.

[0042] like Figure 6As shown, the data acquisition agent collects risk-related data in real time from various data sources, including business systems, sensor networks, external data interfaces, and publicly available information on the internet. The agent employs streaming processing to ensure real-time data transmission. It has multi-protocol adaptability and can handle data of different formats and types. For data from different sources and formats, the agent performs standardization processing to unify data formats and time bases. After initial cleaning and format conversion, the agent semantically annotates the collected data according to the ontology model in the risk knowledge graph. Entity linking technology maps objects mentioned in the data to entities in the knowledge graph. Attribute mapping maps data fields to entity attributes. Relationship recognition extracts relationships between entities from the data. Semantically annotated data is no longer isolated numerical values ​​or text, but rather knowledge graph nodes and edges with clear semantics, which can be directly understood and used by other agents.

[0043] The data analysis agent receives semantically labeled data and performs multi-dimensional in-depth analysis. First, it performs anomaly detection, identifying data points deviating from normal ranges using statistical methods, machine learning models, or rule-based approaches. For time-series data, the agent performs trend analysis, identifying rising, falling, or cyclical patterns. Through correlation analysis, it discovers the correlations and linkages between different risk indicators. The agent also performs aggregation calculations, using the indicator calculation formulas defined in the risk knowledge graph to calculate various risk indicator values ​​from the raw data, including current values ​​and historical trends, providing a quantitative basis for subsequent risk assessment.

[0044] The data analysis agent stores the analysis results in a structured form in the knowledge graph, updates the attribute values ​​of relevant entities, and records any anomalies and trend characteristics discovered during the analysis. Simultaneously, the data analysis agent sends notification messages to other agents, informing them of valuable analytical findings and triggering subsequent processing flows.

[0045] Step S4: Risk Identification and Attribution Analysis Based on Knowledge Graph like Figure 7 As shown, in step S4, when the data analysis agent detects an abnormal signal or risk indicator exceeding a preset threshold, it triggers the risk identification and attribution analysis process.

[0046] The risk identification and attribution analysis agent first identifies risks based on a risk knowledge graph. This agent utilizes the query capabilities of the graph database to retrieve risk patterns matching the current anomaly within the knowledge graph. Through a graph pattern matching algorithm, it identifies the risk type and risk scenario corresponding to the current data.

[0047] The agent also utilizes an ontology reasoning engine for inference analysis. Based on the hierarchical relationships of concepts defined in the knowledge graph, it determines the specific categories of risks through classification reasoning. Based on attribute constraints and rule definitions, it discovers potential risk conflicts and contradictions through consistency checks. Based on causal and correlational relationships, it discovers indirect risk associations and risk transmission paths through transitive reasoning. The reasoning process not only identifies explicit risks but also uncovers hidden potential risks and compound risks.

[0048] Based on risk identification, the agent initiates deep attribution analysis. It extracts the causal relationship network related to the identified risks from the knowledge graph, constructs a risk causal graph, and identifies the root causes and key factors leading to the risks. In the causal graph, risk events are treated as result nodes, various risk factors as cause nodes, and causal relationships as directed edges. The agent employs a causal inference algorithm, combining current data and historical data, to calculate the contribution of each risk factor to the risk event, thus determining the primary and secondary risk sources.

[0049] This intelligent agent not only analyzes direct causes but also traces back to root causes. By performing a reverse traversal of the causal graph, it traces upwards along the causal chain to identify the root causes of risks. For risks resulting from the combined effects of multiple factors, the attribution analysis agent analyzes the interactions and synergistic effects among these factors. The results of the attribution analysis include a list of major risk factors, a ranking of the contributions of each factor, and a description of the causal path. These results provide important evidence for understanding and managing risks.

[0050] The attribution analysis results are not only used for understanding and managing current risks, but can also be fed back into the knowledge graph to enrich causal relationship knowledge. The agent updates the knowledge graph with the analysis results, establishes the association between the current risk event and its causal factors, and records the process and conclusions of the attribution analysis. This newly added knowledge enriches the content of the knowledge graph and provides a reference for the identification and analysis of similar risks in the future.

[0051] Step S5: Multi-dimensional risk assessment and level determination like Figure 8 As shown, in step S5, the risk assessment agent conducts a comprehensive risk assessment based on the results of risk identification and attribution analysis. This agent obtains the standard system for risk assessment from the risk knowledge graph, including assessment dimensions, assessment indicators, weight configurations, and scoring rules. Existing methods such as multi-dimensional weighted assessment and risk matrix grading can be used to measure the severity and scope of impact of risks from different perspectives.

[0052] The risk assessment agent first evaluates the probability dimension of risk, i.e., the likelihood of a risk event occurring. This dimension is assessed based on historical statistical data, the current state of risk factors, and the degree of deviation of risk indicators. By querying historical cases in the knowledge graph, the frequency of risk events under similar conditions is statistically analyzed. Combining the risk factors identified by the risk identification and attribution analysis agent, the probability estimate is adjusted according to the severity of the factors.

[0053] Secondly, the impact dimension of the risk is assessed, namely, the potential losses and scope of impact once the risk event occurs. This assessment needs to consider direct and indirect losses, as well as economic and non-economic impacts. The risk assessment agent retrieves relevant information about the risk subject from the knowledge graph, such as asset size, business scale, and relationships, to assess the degree of impact of the risk on that subject. By analyzing the relationship network in the knowledge graph, the transmission path and scope of the risk are assessed, and other potentially affected subjects are identified.

[0054] The third dimension is the urgency of the risk, that is, the degree of time required to address the risk. This dimension is assessed based on factors such as the rate of change of risk indicators, the evolution trend of risk events, and the available response time window. Trend analysis results provided by data analytics agents play an important role here; a rapidly deteriorating risk trend implies a higher degree of urgency.

[0055] The risk assessment agent integrates the assessment results from multiple dimensions and calculates the overall risk score using a weighted comprehensive scoring method. The weights are obtained from the knowledge graph and can be adjusted according to different risk types and application scenarios. Based on the overall score and the risk level classification criteria defined in the knowledge graph, the risk level is determined, such as low risk, medium risk, high risk, and extremely high risk.

[0056] The risk assessment agent also performs dynamic risk assessments, evaluating not only the current risk status but also predicting future risk evolution trends. By analyzing the historical trajectory and current rate of change of risk indicators, and combining this with knowledge of risk evolution patterns in a knowledge graph, it predicts the future direction and potential severity of the risk. The results of this dynamic assessment provide forward-looking information for early warning decision-making.

[0057] The complete results of the risk assessment include information such as risk level, scores for each dimension, key risk factors, scope of impact, and evolution trend. These results are structured and stored in a knowledge graph and then transmitted to the early warning decision-making agent.

[0058] Step S6: Intelligent Early Warning Decision and Response like Figure 9As shown, in step S6, the early warning decision-making agent makes an early warning decision based on the risk assessment results. This agent retrieves the early warning rule base from the risk knowledge graph and matches corresponding early warning rules according to conditions such as risk level, risk type, and risk subject. The early warning rules define the conditions for triggering an early warning, the criteria for determining the early warning level, and the selection logic for the early warning method.

[0059] The early warning decision-making agent first determines whether an early warning is necessary. For low-risk situations with no deterioration trend, only recording and continuous monitoring may be required, without issuing an early warning. For risks that reach the early warning threshold, an early warning of the corresponding level is required. The determination of the early warning level comprehensively considers factors such as risk level, urgency, and scope of impact. Early warning levels are usually divided into multiple levels, such as attention, warning, alert, and emergency, with different levels corresponding to different response requirements and handling procedures.

[0060] The early warning decision-making agent determines the target audience and notification method based on the characteristics and warning level of the risk. It queries the knowledge graph for information such as the responsible person of the risk entity, relevant management departments, and emergency response teams to determine the recipients of the early warning notification. The notification method is selected based on the warning level and urgency, such as system message, SMS, telephone, or email. Emergency warnings may require multiple methods to be used simultaneously to ensure timely delivery.

[0061] The early warning decision-making agent is also responsible for generating the content of early warning information. This information includes the early warning level, risk type, risk subject, risk description, key indicators, attribution analysis results, impact assessment, and evolution trend. More importantly, the early warning decision-making agent retrieves response measures and contingency plans for this type of risk from a knowledge graph, providing decision support and action guidance to those receiving the warning. The retrieval of response measures is based on conditions such as risk type, risk level, and historical handling experience, matching the most relevant and effective measure suggestions from the knowledge graph.

[0062] For complex risk scenarios, the early warning decision-making agent can also coordinate multiple agents to make joint decisions. For example, when a risk that may trigger a chain reaction is detected, the early warning decision-making agent collaborates with a data analysis agent to simulate the risk's transmission process and scope of impact. It also collaborates with an attribution analysis agent to deeply analyze the root causes and key control points of the risk. Based on the results of these collaborative analyses, a more comprehensive and effective early warning plan is developed.

[0063] After an early warning is generated, the early warning decision-making agent sends the information to the target object through an early warning notification mechanism. Simultaneously, the early warning event is recorded in a knowledge graph, establishing relationships between the early warning event and risk events, risk factors, and countermeasures, providing a data foundation for subsequent effectiveness evaluation and knowledge learning.

[0064] Step S7: Knowledge Learning and Graph Evolution like Figure 10 As shown, in step S7, the learning and optimization agent is responsible for the system's continuous learning and the dynamic evolution of the knowledge graph. This agent continuously collects various types of data and feedback information generated during system operation. This information includes the triggering of warnings, the accuracy assessment of warnings, the actual occurrence of risk events, the effectiveness of countermeasures, and user feedback.

[0065] The learning optimization agent systematically analyzes the collected feedback information. For assessing the accuracy of early warnings, the agent statistically analyzes indicators such as the hit rate, false alarm rate, and missed alarm rate, identifying problems in the early warning rules and evaluation models. For evaluating the effectiveness of countermeasures, the agent analyzes the effectiveness of different measures in different scenarios, summarizing successful experiences and lessons learned. Through comparative analysis, it identifies which risk factors were not accurately identified, which causal relationships were misdefined, and which evaluation weights need adjustment.

[0066] Based on the analysis results, the learning optimization agent optimizes and adjusts various parameters and models of the system. For the early warning threshold, it is dynamically adjusted based on historical data and actual results, seeking the optimal balance between timeliness and accuracy of early warnings. For the weight configuration in the evaluation model, machine learning methods are used to learn the optimal weights from historical data. The calculation formula for risk indicators is modified and improved based on actual conditions.

[0067] The key to learning and optimizing the agent lies in its responsibility for the evolution and updating of the risk knowledge graph. When a new risk type or pattern is discovered, the agent adds new risk concepts and instances to the knowledge graph. When new causal or relational relationships are identified, corresponding edges are added to the knowledge graph. When errors or outdated knowledge are found in the knowledge graph, they are corrected or deleted. The evolution of the knowledge graph is not arbitrary but follows the specifications and constraints of the ontology model to ensure the consistency and integrity of the evolved knowledge graph.

[0068] The learning optimization agent also possesses the ability to learn from new cases. Whenever the system handles a new risk event, the agent stores the complete information of the event as a case in the knowledge graph, including the characteristics of the risk, its causes, its impact, the measures taken, and the results of the handling. These cases become valuable experiential knowledge in the knowledge graph, providing a reference for handling similar risks in the future. Through case-based reasoning technology, the system can retrieve similar historical cases from the knowledge graph when facing new risks, drawing on successful handling experiences.

[0069] The learning optimization agent regularly performs quality checks and optimizations on the knowledge graph. This includes checking for isolated nodes, redundant relationships, inconsistent definitions, and other issues. The knowledge graph's structure is optimized to improve query and reasoning efficiency. The importance of knowledge within the knowledge graph is assessed, high-value knowledge is labeled and protected, and low-value redundant knowledge is removed.

[0070] Through continuous learning and evolution in step S7, the risk knowledge graph is continuously enriched and improved, the system's early warning capability is continuously enhanced, and an intelligent risk early warning system capable of self-learning and self-optimization is formed.

[0071] The method and system of the present invention are further illustrated below with two specific embodiments.

[0072] Example 1: Financial Risk Early Warning Method and System This embodiment applies the present invention to the scenario of credit risk and market risk early warning for financial institutions, and constructs a comprehensive risk early warning method and system based on ontology and multi-agent systems.

[0073] In step S1, a financial risk knowledge graph is constructed. First, a financial risk ontology model is designed, defining core concepts including risk event categories, risk factor categories, risk indicator categories, risk subject categories, and risk level categories. Risk event categories include primary classifications such as credit risk, market risk, liquidity risk, and operational risk. Credit risk is further subdivided into corporate credit risk and personal credit risk, with corporate credit risk further subdivided into default risk, downgrade risk, and concentration risk. Risk factor categories include macroeconomic factors, industry factors, corporate financial factors, market factors, and management factors. Risk indicator categories include dozens of specific indicators such as debt-to-equity ratio, current ratio, quick ratio, interest coverage ratio, market volatility, and default rate.

[0074] The ontology model defines a rich set of semantic relationships. Causal relationships represent the causal connections between risk factors and risk events. For example, a causal chain is defined such that "macroeconomic downturn" leads to "decline in industry prosperity," which in turn leads to "decline in corporate profitability," which leads to "weakened debt repayment ability," and "weakened debt repayment ability" leads to "increased default risk." Association relationships represent the mutual influence between different risks; for example, there is a two-way influence between credit risk and liquidity risk. Impact relationships represent the effect of risk on an entity; for example, a company's default risk affects its credit customers. Indicator relationships represent the mapping between risk and indicators; for example, credit risk is measured through indicators such as the debt-to-equity ratio and the current ratio.

[0075] The ontology model was formally represented using the OWL language, and the knowledge graph data was stored using RDF triples. Historical risk data from the financial institution over the past five years was collected, including financial data, credit data, default records, and market transaction data from thousands of corporate clients. Through entity identification and relation extraction techniques, a risk knowledge graph containing tens of thousands of entity nodes and hundreds of thousands of relation edges was constructed from this data. The knowledge graph also integrates external macroeconomic data, industry data, and market data. Experienced knowledge acquired from risk management experts was transformed into rules and cases within the knowledge graph, forming a rich knowledge foundation.

[0076] In step S2, a multi-agent system is constructed. A data acquisition agent is deployed, connecting to internal data sources such as the core business system, credit management system, and financial system of financial institutions, as well as external data sources such as the People's Bank of China's credit reporting system, industrial and commercial information system, and judicial information system. It also connects to financial market data service providers to obtain real-time market data. The data acquisition agent collects batches of data daily at set times, while simultaneously monitoring key data sources in real time. After cleaning and format conversion, the collected data is semantically annotated according to the ontology model of the risk knowledge graph. Enterprise names are mapped to enterprise entities in the knowledge graph, financial indicators are mapped to entity attributes, and business relationships are mapped to relationship edges between entities.

[0077] The data analysis agent performs multi-dimensional analysis on labeled data. For financial indicators, it calculates the year-on-year and month-on-month change rates of each indicator and identifies abnormal fluctuations. For time series data, it uses methods such as moving averages and exponential smoothing to perform trend analysis and identify rises, falls, or inflection points. For relationships between multiple indicators, it uses correlation and regression analysis to discover the linkages between indicators. The data analysis agent is also responsible for calculating composite indicators, such as calculating a company's overall financial health score based on multiple basic financial indicators.

[0078] The risk identification and attribution analysis agent is specifically responsible for risk attribution analysis. When a company's credit risk indicators show anomalies, the attribution analysis agent extracts a causal network related to that company from a knowledge graph. For example, if a manufacturing company experiences a significant increase in its debt-to-equity ratio, the attribution analysis agent might trace back to the impact of import substitution policies on the company's industry, leading to a decline in overall industry demand. This results in fewer orders and lower revenue for the company, prompting it to increase short-term borrowing to maintain operations, thus increasing its debt-to-equity ratio. The attribution analysis agent not only identifies the direct cause as increased short-term borrowing but also identifies the root causes as changes in industry policies and decreased market demand.

[0079] A risk assessment agent performs a comprehensive evaluation based on the results of data analysis and attribution analysis. The agent retrieves a standard framework for credit risk assessment from a knowledge graph. This framework includes multiple assessment dimensions such as financial status, operational capabilities, industry position, and guarantees, with each dimension containing several specific indicators. The agent calculates scores for each dimension, uses a weighted method to calculate a comprehensive score, and determines the risk level based on the score. For the aforementioned manufacturing company, the comprehensive assessment upgraded its credit risk level from medium to high.

[0080] The early warning decision-making agent makes early warning decisions based on risk assessment results. For enterprises whose risk level rises to high risk, the agent retrieves early warning rules from the knowledge graph to determine whether a level-two early warning needs to be issued. The agent generates early warning information, including basic enterprise information, changes in risk level, key risk indicators, attribution analysis results, and recommended measures. Recommended measures include strengthening post-loan management, increasing monitoring frequency, requiring supplementary collateral, and controlling new credit lines. The early warning information is sent to the enterprise's account manager, risk management department, and relevant leaders via system messages and email.

[0081] The learning-optimized agent continuously collects feedback information from early warnings. For companies that issue warnings, the agent tracks the subsequent situation, recording whether a risk event actually occurred and the effectiveness of the countermeasures. Through statistical analysis, the agent identifies that companies in certain industries are particularly sensitive to changes in macroeconomic policies and adjusts the warning thresholds for these industries accordingly. The agent also learns from new risk cases, storing successful warnings and successful responses as typical examples in a knowledge graph, enriching its content.

[0082] During steps S3 to S7, multiple intelligent agents collaborate to form a complete risk warning process. The data acquisition agent collects the latest data from thousands of companies daily and performs semantic annotation. The data analysis agent analyzes this data and identifies dozens of companies with abnormal risk indicators. The attribution analysis agent conducts in-depth attribution analysis on these abnormal companies to identify the causes of the risks. The risk assessment agent, after comprehensive evaluation, determines that the risk levels of more than ten companies need adjustment. The warning decision agent issues warnings for companies with rising risk levels and lifts warnings for companies with declining risk levels. The learning and optimization agent continuously learns and optimizes, constantly improving the accuracy of the system's warnings.

[0083] After six months of operation, the system successfully issued early warnings for multiple potential credit risk events, achieving an accuracy rate of over 85%. Through timely warnings and corresponding countermeasures, potential losses of tens of millions of yuan were avoided. The system's risk knowledge graph has also grown from tens of thousands of entities initially to over one hundred thousand entities, significantly improving the richness and accuracy of its knowledge.

[0084] Example 2: Industrial Production Safety Risk Early Warning Method and System This embodiment applies the invention to a production safety risk early warning scenario in a large chemical enterprise. This enterprise has multiple production workshops and hundreds of key pieces of equipment, and faces various safety risks such as fire, explosion, and leakage.

[0085] In step S1, an industrial safety risk knowledge graph is constructed. The designed ontology model includes core concepts such as safety risk events, risk sources, safety indicators, protective measures, and emergency plans. Safety risk events include types such as fire, explosion, toxic gas leaks, mechanical injuries, and high-temperature burns. Risk sources include hazardous chemicals, high-temperature and high-pressure equipment, special equipment, and electrical equipment. Safety indicators include physical quantities such as temperature, pressure, liquid level, flow rate, gas concentration, vibration, and noise. The ontology model defines the triggering relationship between risk sources and risk events, the correlation between safety indicators and risks, and the chain relationships between risk events.

[0086] Safety production data from the company over the past ten years was collected, including historical accident records, hazard investigation records, equipment maintenance records, and emergency drill records. A wealth of safety knowledge was extracted from this data, and a knowledge graph was constructed containing thousands of risk source entities, hundreds of risk scenarios, and tens of thousands of monitoring data points. The knowledge graph also integrates external knowledge such as safety standards, regulations, and best practices in the chemical industry.

[0087] In step S2, a multi-agent system is constructed. Data acquisition agents are deployed in each production workshop. These agents are connected to hundreds of IoT sensors within the workshop, such as temperature sensors, pressure sensors, gas detectors, level gauges, and flow meters, to collect equipment operation data and environmental monitoring data in real time. The data acquisition agents collect key parameters at a rate of seconds, perform preliminary anomaly detection on the collected data, and immediately perform semantic annotation and trigger alarms for data that exceeds the normal range.

[0088] The data analysis agent comprehensively analyzes data from multiple sensors. For temperature data, it not only monitors whether the absolute value exceeds the limit but also analyzes the rate and trend of temperature change. For pressure data, it analyzes pressure fluctuation characteristics and identifies abnormal pressure pulsations. For multiple related parameters, it analyzes the correlation between them; for example, a rise in temperature is usually accompanied by a rise in pressure, and an anomaly where temperature rises but pressure remains unchanged may indicate equipment failure. The data analysis agent also establishes a baseline model for the equipment's normal operation and identifies anomalies by the degree of deviation from the baseline.

[0089] The risk identification and attribution analysis agent identifies and analyzes the causes of safety risks when it detects a safety risk signal. For example, when the temperature of a reactor rises abnormally, the agent retrieves possible causes from the knowledge graph, including cooling system malfunction, reaction runaway, stirring system malfunction, and temperature sensor malfunction. The agent further examines relevant information such as cooling water flow data, stirring motor current data, and other temperature sensor data, using a process of elimination and causal reasoning to determine the most likely cause. In one real-world case, the agent's analysis revealed that the root cause of the temperature rise was the accidental closure of a cooling water pipe valve, rather than a problem with the reaction itself.

[0090] A risk assessment agent evaluates the severity and urgency of safety risks. Regarding the aforementioned abnormal temperature in the reactor, the agent retrieves the reactor's safety parameters from a knowledge graph, including normal operating temperature, maximum permissible temperature, and hazardous temperature. Based on factors such as the current temperature, rate of temperature rise, and margin before the hazardous temperature, the agent assesses the severity of the risk. It also assesses the risk's impact range; the reactor is located in the center of the workshop, surrounded by other equipment and storage tanks, and an accident could trigger a chain reaction. The comprehensive assessment determines this risk to be high, requiring immediate action.

[0091] The early warning decision-making agent makes rapid decisions based on risk assessment results. For high-risk situations, the agent immediately triggers audible and visual alarms, notifies workshop personnel via the broadcast system, and simultaneously sends emergency notifications to the workshop director, safety management department, and emergency command center. The agent retrieves emergency response plans from a knowledge graph; for abnormal reactor temperatures, the plan includes immediately stopping feeding, initiating emergency cooling, inspecting the cooling system, and preparing emergency supplies. The warning information includes detailed operational instructions to help on-site personnel respond quickly.

[0092] The learning-optimizing agent learns from each warning event. For the aforementioned abnormal reactor temperature event, post-event analysis revealed that human error caused the valve to close. The learning-optimizing agent added this case to its knowledge graph and updated its attribution analysis knowledge base, adding a rule that "when cooling water flow suddenly drops and there are no equipment fault alarms, the valve status should be checked." The learning-optimizing agent also recommended adding status monitoring sensors to critical valves to fundamentally prevent similar problems.

[0093] During the operation of the multi-agent system, close collaboration was established among the agents. In one instance, while analyzing gas concentration data in a storage tank area, the data analysis agent detected a slow upward trend in the concentration of combustible gas. Although the concentration had not yet reached the alarm threshold, the trend was abnormal. The data analysis agent notified the attribution analysis agent and the risk assessment agent of this finding. The attribution analysis agent analyzed possible leak sources, while the risk assessment agent assessed the time it would take for the concentration to reach the lower explosive limit if it continued to rise. After comprehensive judgment, the early warning decision-making agent issued a preventative warning, requesting on-site personnel to conduct an investigation. Ultimately, a minor leak was found at a pipeline flange, which was promptly addressed, preventing a potentially major accident.

[0094] In its year of operation, the system has successfully issued warnings of dozens of potential safety risks, including several high-risk situations that could have led to major accidents if not addressed promptly. The system's false alarm rate is kept below 5%, earning high praise from the company's safety management department. Through continuous learning and optimization, the system's risk identification capabilities have been continuously improved, and the safety knowledge in its knowledge graph has become increasingly rich and accurate.

[0095] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Those skilled in the art can modify or make equivalent substitutions to the technical solutions of the present invention. The scope of protection of the present invention is defined by the claims.

Claims

1. A risk warning method based on ontology and multi-agent systems, comprising the following steps: Construct or acquire a risk knowledge graph based on an ontology model; Risk data is collected by an intelligent agent connected to a risk knowledge graph, and semantic annotation is performed based on the ontology model. Anomaly detection is performed on the labeled risk data by another intelligent agent connected to the risk knowledge graph, and risk indicator analysis is performed based on the risk knowledge graph. When an anomaly or risk indicator is detected that exceeds a set threshold, another intelligent agent connected to the risk knowledge graph performs risk identification and attribution analysis based on the risk knowledge graph. Risk assessment is conducted by another agent connected to the risk knowledge graph, based on the risk identification results and attribution analysis results, according to the risk knowledge graph's evaluation criteria. Another agent connected to the risk knowledge graph performs early warning processing based on the risk assessment results and the early warning rules of the risk knowledge graph.

2. The method according to claim 1, characterized in that, The system learns feedback information through another agent connected to the risk knowledge graph to optimize the early warning method and update the risk knowledge graph.

3. The method according to claim 1, characterized in that, By analyzing the characteristics and needs of the risk domain, risk concepts are identified, and the attribute features of the concepts and the semantic relationships between the concepts are defined to construct the ontology model. Based on the ontology model, multi-source heterogeneous risk information, including historical risk data, expert knowledge, and business rules, is transformed into entities and relationships in a knowledge graph. Risk knowledge is extracted through entity recognition, relationship extraction, and attribute filling methods, and organized and stored according to the ontology model specifications to form a risk knowledge graph.

4. The method according to claim 1, characterized in that, Risk identification also involves analyzing the relationships within the risk knowledge graph, assessing the transmission path and scope of the risk, and identifying other entities that may be affected.

5. The method according to claim 1, characterized in that, The early warning process also includes obtaining risk response measures from a risk knowledge graph.

6. The method according to claim 1, characterized in that, The risk assessment includes a risk assessment at the current moment, as well as a prediction of the future evolution trend of the risk.

7. The method according to claims 1-6, characterized in that, The agents communicate with each other through message passing, and the message content is represented using the concepts and relationships defined in the risk knowledge graph.

8. A risk warning system based on ontology and multi-agent systems, including... A risk knowledge graph constructed based on an ontology model; A multi-agent system connected to the risk knowledge graph includes A data acquisition agent collects risk data and performs semantic annotation based on the ontology model. The data analysis agent performs anomaly detection on the risk data labeled by the data collection agent and performs risk indicator analysis based on the risk knowledge graph definition. The risk identification and attribution analysis agent performs risk identification and attribution analysis based on the risk knowledge graph when the data analysis agent detects anomalies or risk indicators exceed a set threshold. The risk assessment agent performs risk assessments based on the risk identification and attribution analysis results of the risk identification and attribution analysis agent, according to the assessment criteria of the risk knowledge graph. The early warning decision-making agent performs early warning processing according to the early warning rules of the risk knowledge graph based on the risk assessment results of the risk assessment agent. The learning agent optimizes itself by learning from feedback information to improve early warning methods and update the risk knowledge graph.

9. An electronic device, characterized in that, The electronic device includes: a processor and a memory storing computer program instructions; the processor, when executing the computer program instructions, implements the steps of the method as described in claims 1-7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer program instructions that, when executed by a processor, implement the steps of the method as described in claims 1-7.

Images