Language model collaborative target event processing method and device, equipment and medium

Abstract

The invention relates to the technical field of artificial intelligence, can be applied to business scenes such as financial science and technology and medical health, and discloses a target event processing method, device, equipment and medium based on language model collaboration.The method comprises the steps that a target event is monitored and recognized, multi-source data associated with the target event are collected, the multi-source data are fused to generate fused data, and the fused data are sent to a server; and inputting the fusion data into a pre-training language model to generate a decision scheme, performing decision coordination and matching based on the decision scheme, distributing a processing agent, executing a matched decision task by using the processing agent, and generating a decision result. According to the invention, through fusion of multi-source data and pre-training language model deep reasoning and combination of an intelligent agent dynamic cooperation mechanism, intelligent perception, rapid response and cooperative processing of complex events are realized, and the adaptive ability and processing efficiency of the system in response to unknown faults and complex tasks are improved.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, and in particular to a method, apparatus, device, and storage medium for target event processing using language model collaboration. Background Technology

[0002] In the fintech sector, as business systems become increasingly complex, the requirements for real-time performance, stability, and security in financial services are also rising. Traditional emergency response solutions largely rely on manually pre-set rule engines or automated operation and maintenance scripts to handle system failures, network anomalies, and data security risks. However, these solutions are highly dependent on fixed rules and manual configuration. When faced with new financial fraud methods, sudden system anomalies, or complex combinations of failures, their response is slow and inflexible, failing to quickly and accurately coordinate and handle multiple tasks, severely impacting business continuity and system stability.

[0003] In the healthcare sector, the integration of hospital information systems, telemedicine platforms, and smart devices is constantly increasing. Medical data comes from a wide range of sources and has a complex structure, making stable system operation crucial for patient safety. Traditional medical information systems often employ fault handling solutions based on single models or fixed scripts, lacking intelligent reasoning and multi-source information fusion capabilities. Especially when facing complex equipment malfunctions, system performance bottlenecks, or sudden public health emergencies, they struggle to coordinate multiple tasks and dynamically manage situations in a timely and efficient manner, potentially leading to interruptions in healthcare services, data security risks, or impacting the quality of treatment decisions.

[0004] In existing intelligent system solutions, some systems attempt to introduce monolithic AI models such as decision trees and traditional machine learning models to assist in decision-making. However, these models have limited reasoning capabilities and lack adaptability to dynamic and complex scenarios. Especially in applications requiring multi-task collaboration, dynamic allocation of agents, or multi-source data fusion analysis, they struggle to meet the dual demands of real-time performance and accuracy. Meanwhile, while current mainstream automated operation and maintenance tools possess some task automation capabilities, they generally lack cross-system, multi-task collaboration, and intelligent reasoning functions. They cannot cope with complex and dynamic business needs, particularly when system anomalies are frequent or new threats emerge. In these situations, manual intervention is still required for decision-making and resource allocation, indicating that the system's adaptability and intelligence level need improvement. Summary of the Invention

[0005] The main objective of this invention is to provide a method, apparatus, device, and storage medium for target event processing in collaboration with language models, aiming to solve the technical problems in the prior art where preset rule engines and single-entity reasoning models are difficult to adaptively handle unknown faults, and automated tools lack multi-agent collaboration and deep reasoning capabilities.

[0006] To achieve the above objectives, the present invention provides a target event processing method for language model collaboration, comprising:

[0007] Monitor and identify target events, and collect multi-source data associated with the target events;

[0008] The multi-source data is fused to generate fused data;

[0009] The fused data is input into a pre-trained language model, and a decision scheme is generated through the pre-trained language model.

[0010] Decision coordination is performed according to the decision scheme, and processing agents are matched and assigned to the decision tasks in the decision scheme.

[0011] The processing agent executes the matching decision-making task and generates decision results.

[0012] Furthermore, to achieve the above objectives, the present invention provides a target event processing device for language model collaboration, comprising:

[0013] The event monitoring module is used to monitor and identify target events and collect multi-source data associated with the target events;

[0014] The data fusion module is used to fuse the multi-source data to generate fused data;

[0015] The language model inference module is used to input the fused data into a pre-trained language model and generate a decision scheme through the pre-trained language model.

[0016] The agent coordination module is used to coordinate decisions according to the decision scheme, and to match and assign processing agents for the decision tasks in the decision scheme.

[0017] The task execution module is used to execute the matched decision-making task through the processing agent and generate decision results.

[0018] Furthermore, to achieve the above objectives, the present invention also provides a computer device, the computer device including a memory, a processor, and a language model collaboration target event handling program stored in the memory and executable on the processor, wherein when the language model collaboration target event handling program is executed by the processor, it implements the steps of the language model collaboration target event handling method as described above.

[0019] Furthermore, to achieve the above objectives, the present invention also provides a computer-readable storage medium storing a target event handling program for language model collaboration, wherein when the target event handling program for language model collaboration is executed by a processor, it implements the steps of the target event handling method for language model collaboration as described above.

[0020] Beneficial Effects: This invention relates to the field of artificial intelligence technology and can be applied to business scenarios such as fintech and healthcare. It discloses a method, apparatus, device, and medium for target event processing using language model collaboration, comprising: monitoring and identifying target events; collecting multi-source data associated with the target events; fusing the multi-source data to generate fused data; inputting the fused data into a pre-trained language model to generate a decision scheme; coordinating decisions based on the decision scheme; matching and assigning processing agents; utilizing the processing agents to execute matched decision tasks; and generating decision results. This invention, by fusing multi-source data with deep inference from a pre-trained language model and combining it with a dynamic agent collaboration mechanism, achieves intelligent perception, rapid response, and collaborative processing of complex events, improving the system's adaptability and processing efficiency when dealing with unknown faults and complex tasks. Attached Figure Description

[0021] The present invention will be further described below with reference to the accompanying drawings and embodiments. In the accompanying drawings:

[0022] Figure 1 This is a schematic diagram of an application environment for a language model collaborative target event processing method according to an embodiment of the present invention;

[0023] Figure 2 This is a flowchart illustrating an embodiment of the target event processing method for language model collaboration of the present invention;

[0024] Figure 3 This is a schematic diagram of the functional modules of a preferred embodiment of the target event processing device for language model collaboration of the present invention;

[0025] Figure 4 This is a schematic diagram of the structure of a computer device according to an embodiment of the present invention;

[0026] Figure 5 This is another structural schematic diagram of a computer device according to one embodiment of the present invention. Detailed Implementation

[0027] It should be understood that the specific embodiments described herein are for illustrative purposes only and are not intended to limit the scope of the invention.

[0028] The target event processing method for language model collaboration provided in this invention can be applied to, for example... Figure 1In this application environment, the user terminal communicates with the server via a network. The server can monitor and identify target events through the user terminal, collect multi-source data associated with the target events, fuse the multi-source data to generate fused data, input the fused data into a pre-trained language model to generate a decision scheme, coordinate decisions based on the decision scheme, match and assign processing agents, utilize the processing agents to execute the matched decision tasks, and generate decision results. This invention achieves intelligent perception, rapid response, and collaborative processing of complex events by fusing multi-source data with deep inference of a pre-trained language model and combining it with a dynamic agent collaboration mechanism, thereby improving the system's adaptability and processing efficiency when dealing with unknown faults and complex tasks. The user terminal can be, but is not limited to, various personal computers, laptops, smartphones, tablets, and portable wearable devices. The server can be implemented using a standalone server or a server cluster consisting of multiple servers. The following detailed description of specific embodiments further illustrates this invention.

[0029] Please see Figure 2 , Figure 2 This is a flowchart illustrating an embodiment of the language model collaboration target event processing method provided by the present invention. It should be noted that although the logical order is shown in the flowchart, in some cases, the steps shown or described may be performed in a different order than that shown here.

[0030] like Figure 2 As shown, the target event processing method for language model collaboration proposed in this invention includes the following steps:

[0031] S10, Monitor and identify target events, and collect multi-source data associated with the target events;

[0032] In this embodiment, during the monitoring and identification of target events and the collection of multi-source data associated with those events, a dynamic perception and intelligent event identification mechanism is employed to address uncertainties and sudden risks encountered by the system in actual operation, thereby improving the overall system's environmental adaptability and fault response level. Monitoring actions are based on the overall system operating status, encompassing multi-dimensional monitoring information sources including hardware devices, software modules, network resources, and external environmental factors. These sources specifically include, but are not limited to, sensor data, log information, network traffic, user operation behavior, and peripheral system interaction information. Operating status information can be obtained by deploying distributed acquisition nodes, edge detection modules, or internal system embeddings. During the acquisition process, the continuity of real-time data streams and the integrity of the system structure are combined to ensure that monitoring information covers the entire system. To avoid data silos, while acquiring information, structured tagging, time synchronization mechanisms, and data source mapping relationships are used to ensure that the monitoring information has temporal consistency and spatial correlation.

[0033] Event identification, based on acquired system operational status information, employs feature extraction and pattern discrimination algorithms to comprehensively analyze key indicators, characteristic patterns, and abnormal change trends within the operational status information. This identifies potential anomalies, combinations of anomalies, or complex fault scenarios caused by internal or external factors. During event identification, a multi-model fusion mechanism based on rule-based reasoning, statistical learning, and deep learning can be introduced to perform multi-dimensional discrimination and aggregation classification of different types of risk sources, such as abnormal behavior, performance mutations, security threats, and operational bottlenecks. Target events refer to specific abnormal situations requiring key system response or further handling. They typically possess clear triggering conditions, impact scope, and risk levels, and their sources include, but are not limited to, operational anomalies, performance degradation, security vulnerabilities, external attacks, or environmental changes.

[0034] Multi-source data acquisition is based on identified target events, dynamically determining the set of multi-source data associated with those events. Multi-source data includes structured, unstructured, and semi-structured data, with specific sources including, but not limited to, system logs, device status information, application performance metrics, network topology information, historical fault data, user behavior records, business processing data, and environmental monitoring information. During data acquisition, a layered distributed data acquisition architecture is adopted, combining data type and structure, supporting a flexible combination of synchronous data streams and asynchronous data snapshots to achieve large-scale, low-latency, and high-accuracy data acquisition. Data sampling strategies, intelligent compression algorithms, and data deduplication mechanisms can be used during acquisition to reduce data redundancy and improve system resource utilization efficiency.

[0035] After data from different sources is collected, it needs to undergo unified format conversion and data identification standards to ensure that the data content is comparable, analyzable, and correlated. In the process of acquiring and managing multi-source data, data tagging systems, metadata management mechanisms, and data lineage tracing methods can be combined to establish cross-source data correlation information, thereby improving the effective utilization rate of data in subsequent analysis, fusion, and reasoning processes.

[0036] For industrial production environments, edge monitoring nodes and a central data aggregation platform are deployed to acquire real-time equipment status parameters, production process indicators, and environmental monitoring data. Combined with industrial control system logs and historical fault records, this enables comprehensive monitoring of system operation status and anomaly risks. In intelligent transportation scenarios, roadside sensing units, vehicle-mounted sensors, and traffic management system data are used to dynamically perceive traffic flow, abnormal congestion, and accident information, achieving efficient identification of target events and acquisition of multi-source data. In the field of large data center operation and maintenance, distributed monitoring probes and system log collection tools are used to monitor server status, network traffic, application performance, and security events in real time, quickly identifying target anomalies, collecting related data, and supporting subsequent intelligent response and decision optimization.

[0037] In healthcare data processing scenarios, by combining electronic medical record systems, medical equipment monitoring information, patient physiological parameters, and medical behavior logs, abnormal risks in the medical process can be dynamically monitored, such as equipment failure, diagnostic deviations, and abnormal health indicators. Based on event identification results, multi-source data, including historical medical records, equipment status, medical records, and medical staff operation information, can be collected to provide data support for medical process optimization and risk control. In fintech business processing, by combining transaction system logs, account information changes, fund flow monitoring, and risk model outputs, abnormal situations in transaction systems and business processes can be monitored in real time. Target events involving fraud, compliance risks, and system failures can be identified, and multi-source data, including transaction records, account attributes, behavioral trajectories, and external data source information, can be dynamically collected to support financial business risk warnings and system adaptive adjustments.

[0038] This embodiment, through system-level dynamic monitoring and intelligent event recognition, can overcome the limitations of traditional reliance on static rules and manual intervention, and achieve efficient perception and identification of complex, sudden, and cross-system risk sources. Combined with comprehensive collection and structured management of multi-source data, it breaks down information silos, improves the coverage, timeliness, and effectiveness of data acquisition, and provides a high-quality data foundation for subsequent data fusion, intelligent reasoning, and multi-agent collaborative decision-making, effectively enhancing the system's adaptability and intelligent response level in the face of complex dynamic environments.

[0039] S20, perform fusion processing on the multi-source data to generate fused data;

[0040] In this embodiment, multi-source data refers to a collection of data originating from different systems, devices, applications, or information channels. This includes, but is not limited to, structured, semi-structured, and unstructured data, specifically data collected by sensors, log data, data output from business systems, network traffic data, user operation data, environmental monitoring data, and other related information. Multi-source data differs in terms of source, format, timeliness, granularity, accuracy, encoding standards, and information integrity. To ensure comparability, computability, and consistency of data from different sources, a data fusion process must be performed on the multi-source data.

[0041] Fusion processing refers to the process of transforming heterogeneous, structurally inconsistent, and informationally scattered multi-source data into a dataset that meets the requirements of unified logical expression through a series of actions including data standardization, structural unification, semantic alignment, redundancy removal, data completion, relationship mapping, time synchronization, and information aggregation, thus generating fused data. Specifically, this includes: First, using data cleaning rules to identify and remove missing values, duplicate values, invalid data, and noisy data from multi-source data to ensure data quality. Second, converting the cleaned data according to a unified data format standard to ensure consistency in encoding methods, field types, unit expressions, and time formats across different sources. Then, performing semantic alignment operations on key fields in the data based on a field semantic mapping table or data dictionary to ensure consistent expression of the same entities, attributes, or events across different data sources. Further, reconstructing the logical relationships in the multi-source data based on entity relationships, timestamp information, and spatial location information to form a topological structure or multi-dimensional data network. Finally, integrating structurally unified, semantically aligned data with logical relationships into fused data through data aggregation, computational extraction, and feature combination.

[0042] Fusion data refers to a data set generated based on the fusion processing of multi-source data, possessing a unified data structure, unified coding standards, semantic consistency, and complete logical relationships. Fusion data supports unified processing, deep computing, intelligent reasoning, and cross-system collaborative operations in downstream systems. Fusion data can be expressed using multidimensional arrays, structured tables, graph structures, nested data objects, or other suitable standard data structures, with the specific form determined according to the requirements of the subsequent processing system.

[0043] In the healthcare sector, multi-source data includes physiological parameter monitoring data from wearable devices, electronic health records from hospital information systems, real-time examination results from clinical testing equipment, medication usage records, abnormal data from disease early warning systems, and air quality data from environmental monitoring systems. Through fusion processing, data output from different medical devices, systems, and platforms can be transformed into a unified health management data structure. This enables a comprehensive expression of patient vital signs, health risks, medical history, and environmental factors, generating fused data that can be used by intelligent decision-making systems or auxiliary diagnostic models.

[0044] In the fintech business, multi-source data includes transaction logs from transaction systems, user terminal behavior logs, transaction request information from payment gateways, anomaly detection data from risk control systems, and credit data and public opinion information provided by external data service providers. Through fusion processing, heterogeneous data from transaction links, user interactions, and external data sources are integrated into a unified customer risk profile, transaction behavior pattern, and business data set, generating fused data that provides standardized input for intelligent risk control systems, anti-fraud models, and automated decision-making systems.

[0045] The implementation method of fusion processing can be adjusted according to business needs, data types, and differences in system architecture, specifically including:

[0046] It can perform distributed cleaning, transformation and fusion of multi-source data based on a distributed data processing framework, and is suitable for high-concurrency and massive data scenarios;

[0047] Alternatively, edge computing devices can perform partial fusion processing on the data source side, reducing data transmission pressure and central computing load;

[0048] It can also be combined with a data platform, a unified data model, or a semantic tagging system to improve the semantic alignment efficiency and structural uniformity of multi-source data;

[0049] Furthermore, for unstructured text, images, audio, video, and other information in multi-source data, technologies such as natural language processing, image recognition, and audio analysis can be introduced to assist in the extraction and fusion of structures.

[0050] This embodiment effectively eliminates data structure differences, inconsistent coding standards, and inconsistent semantic expressions between different data sources through multi-source data fusion processing, thereby improving data quality and enhancing information integrity and logical coherence. The fused data provides downstream systems with a high-quality, standardized, and structurally unified input data foundation, ensuring the accuracy and efficiency of subsequent intelligent reasoning, decision analysis, task allocation, and system collaborative operations, and enabling multi-system data interoperability and collaborative processing in complex business environments.

[0051] S30, input the fused data into the pre-trained language model, and generate a decision scheme through the pre-trained language model;

[0052] In this embodiment, fused data is a data set formed by constructing multi-source data through structural unification, semantic standardization, and logical association. It possesses the characteristics of structural integrity, consistent encoding, clear semantics, and logical closure. Fusion data provides standardized data input for downstream intelligent computing and reasoning analysis. Fusion data may include structured data tables, vector data, graph structured data, text data, or combinations of multimodal data. The specific content is determined based on business needs and the results of upstream data fusion processing.

[0053] Pre-trained language models are artificial intelligence models trained on large-scale datasets through self-supervised or unsupervised training mechanisms. They possess broad capabilities in language understanding, semantic modeling, and knowledge reasoning. Common model architectures include large-scale language models based on the Transformer architecture, which feature multi-layer neural network structures, positional encoding mechanisms, and multi-head self-attention mechanisms. These models are able to capture complex semantic relationships and contextual logic in the input data. By training on general corpora, domain-specific data, or task-related data, pre-trained language models acquire rich semantic information, knowledge associations, and reasoning patterns, and can serve as foundational models or core reasoning tools for various intelligent tasks.

[0054] Inputting fused data into a pre-trained language model refers to performing data preprocessing, format conversion, and input structure construction operations according to the model's input format and data specifications to ensure that the fused data meets the model's input requirements. Specifically, this includes: for structured fused data, transforming the data into a parsable input vector or embedding representation through field selection, feature encoding, numerical standardization, and vector construction; for text-based fused data, constructing a sequence structure conforming to the language model's input format using word segmentation, encoding, positional embedding, and special tagging; and for graph-structured fused data, mapping the data into an input form that supports structural reasoning based on graph neural networks or structural encoding methods.

[0055] Generating decision-making solutions through pre-trained language models refers to leveraging the model's semantic understanding, pattern recognition, and logical reasoning capabilities to perform reasoning, computation, and analysis operations based on fused input data, outputting decision-making solutions tailored to actual business needs. These decision-making solutions can include multi-level and multi-dimensional intelligent reasoning results, specifically in forms including but not limited to task allocation plans, disposal strategy suggestions, response priority determination, resource coordination schemes, or other executable decision-making information for complex business systems.

[0056] During inference, the pre-trained language model combines its internal parameter knowledge, semantic structure, and inference mechanism to analyze entity elements, logical relationships, and contextual information in the fused data, uncover hidden connections, identify key features, and derive the optimal or suitable decision output. The decision scheme is characterized by clear structure, logical rationality, strong interpretability, and feasibility, meeting the intelligent decision-making requirements of the system.

[0057] In the healthcare sector, integrated data can include patient vital signs data, disease diagnosis information, laboratory test results, and medical resource data. By inputting this integrated data into a pre-trained language model, decision-making solutions for medical emergency response can be generated. These solutions include treatment pathway recommendations, patient triage suggestions, medical resource coordination, and emergency response decisions, thereby improving emergency response efficiency and the utilization of medical resources.

[0058] In the fintech business, integrated data can include transaction records, risk indicators, user behavior data, and external credit information. By inputting the integrated data into a pre-trained language model, decision-making solutions for risk prevention and anomaly response are generated. These solutions include transaction freeze recommendations, risk level assessments, fund path restructuring, and system linkage strategies, enabling real-time intelligent risk control and multi-system collaborative responses.

[0059] In different business scenarios, the choice of pre-trained language models can be adjusted according to task complexity, data characteristics, and inference needs. Large-scale general-purpose language models can be used to enhance complex semantic reasoning capabilities, domain-specific fine-tuned models can be deployed to improve adaptability to specific tasks, and hybrid multi-model architectures can be used to achieve comprehensive reasoning and multi-task collaboration capabilities.

[0060] This embodiment inputs fused data into a pre-trained language model. Based on the model's deep semantic understanding, complex logical reasoning, and knowledge representation capabilities, it can efficiently and accurately generate intelligent decision-making solutions tailored to actual business needs. Compared to traditional methods that rely on preset rules or single models, this process has advantages such as strong adaptability, high reasoning ability, fast decision-making efficiency, and strong multi-task collaboration capabilities. It overcomes the response bottlenecks in existing technologies when dealing with unknown faults, complex environments, and dynamic tasks, significantly improving the overall intelligent decision-making level and business collaboration efficiency of the system.

[0061] S40, perform decision coordination according to the decision scheme, and match and assign processing agents to the decision tasks in the decision scheme;

[0062] In this embodiment, the decision-making scheme is a structured or semi-structured output generated from fused data through a pre-trained language model. It possesses clear task content, priority information, and processing logic, guiding subsequent task collaboration and agent execution. The decision-making scheme may include a task list, task dependencies, resource requirements, execution conditions, and policy rules. The task list specifies the concrete tasks to be processed; task dependencies reflect the logical order and correlation between tasks; resource requirements describe the computational, communication, or external environmental conditions required for task execution; execution conditions clarify the criteria for task triggering or completion; and policy rules provide decision constraints and optimization directions.

[0063] Decision coordination based on the decision scheme refers to integrating system resources, analyzing task relationships, and evaluating the execution environment based on the specific content of the decision scheme, to formulate a reasonable task scheduling plan and resource allocation scheme, ensuring orderly task collaboration and efficient resource utilization in a multi-task, high-concurrency environment. Decision coordination includes task parsing, dependency analysis, resource evaluation, and collaborative optimization processes. Task parsing extracts task elements and execution parameters from the decision scheme using structured methods; dependency analysis identifies the sequence relationships and data transmission logic between tasks; resource evaluation obtains available resources and performance indicators based on system state and processing agent information; and collaborative optimization combines task requirements and resource status to formulate the optimal task matching and agent allocation scheme.

[0064] Matching and assigning processing agents to decision-making tasks within a decision-making scheme refers to the process of selecting agents, binding tasks, and assigning them based on the analysis results of decision coordination and the capabilities, current load, geographical location, and historical performance of each processing agent. Processing agents are hardware and software entities with independent execution capabilities, resource management functions, and status feedback mechanisms. They can include virtual computing units, physical devices, software agents, edge nodes, or other system components with autonomous decision-making and task execution capabilities. Task matching ensures a high degree of matching between tasks and agents in terms of functionality, performance, and environmental requirements through feature comparison, capability adaptation, and policy constraints. Agent allocation is based on the principles of load balancing, maximizing resource utilization, and optimizing response efficiency. Tasks are dynamically assigned to suitable processing agents, and task information is sent through control signals or communication interfaces to drive the agents into the task execution state.

[0065] In the field of healthcare, decision-making solutions can include patient management tasks and medical resource allocation strategies in the event of a public health emergency. During the decision-making and coordination process, various medical institutions, emergency resources and personnel capabilities are combined to dynamically match medical units with professional capabilities, available resources and geographical advantages as processing agents, so as to achieve efficient patient transfer, on-site treatment and remote collaboration.

[0066] In the fintech business, decision-making solutions can include tasks for isolating and handling cross-system financial risk events and account security control strategies. During the decision-making and coordination process, information from payment systems, risk control platforms and regulatory interfaces is integrated, and intelligent risk control modules with security authentication, real-time monitoring and multi-system linkage capabilities are dynamically matched as processing agents to quickly respond to potential risks and execute account freezing or transaction blocking operations.

[0067] In different scenarios, the type and allocation logic of the processing agents can be adjusted according to the system architecture, task complexity and response requirements. A centralized agent cluster can be used to improve the high-density task processing capability, a distributed edge agent can be used to enhance the on-site response speed, and a multi-level collaborative architecture can be used to achieve integrated task collaboration and resource optimization of cloud, edge and terminal.

[0068] This embodiment coordinates decisions based on a decision-making scheme, dynamically matching and allocating suitable processing agents by combining system task relationships, resource status, and processing agent capabilities. This enables efficient and flexible agent collaboration and resource optimization in complex task environments. This process overcomes the limitations of traditional fixed rules and static allocation models in dealing with unknown faults, multi-task conflicts, and resource bottlenecks, significantly improving the overall system's task response efficiency, agent utilization, and collaborative decision-making level in complex environments.

[0069] S50, the processing agent executes the matching decision task and generates a decision result.

[0070] In this embodiment, the processing agent is a software or hardware entity or system unit with autonomous task execution capabilities, which may include virtual machines, container instances, robot units, edge devices, intelligent control modules, or other computing resources, control nodes, or execution terminals capable of independently completing task allocation. The matched decision task refers to a specific task selected through a decision coordination process that is compatible with the corresponding processing agent's functions, performance, and state, based on a combination of decision-making schemes, task content, and agent capabilities. The matching process ensures compatibility and efficient collaboration between the task and the agent in terms of functional requirements, computing resources, geographical location, and time requirements.

[0071] Execution matching decision tasks refers to the entire process by which a processing agent, after receiving a task instruction, independently or collaboratively completes the target task based on task parameters, policy constraints, and the current system state. This includes multiple stages: task initiation, task process control, data acquisition and feedback, status monitoring, anomaly handling, and result generation. Task instructions are structured or semi-structured data sets that clearly define the task objectives, operational steps, input / output requirements, and execution standards. After parsing the instructions, the processing agent automatically enters the task execution flow.

[0072] During task execution, the processing agent, based on the task content and system environment, schedules local resources or calls external interfaces to perform operations such as data processing, environmental perception, action control, and communication interaction. It monitors its own status and task progress in real time, and promptly activates redundancy mechanisms or sends out early warning information upon detecting anomalies. In some scenarios, task execution also involves multi-agent collaboration, multi-task concurrency, or dynamic task adjustment. The processing agent dynamically adjusts its operational strategies based on task dependencies and system feedback to ensure the overall task is completed efficiently and reliably.

[0073] Generating decision results refers to the process by which an intelligent agent, after completing a decision-making task, outputs structured task result information based on task execution data, system feedback information, and operational effects. This information includes task status, execution metrics, anomaly information, and performance evaluation. Decision results can serve as direct feedback on task completion and also provide data support for subsequent system learning, agent optimization, and overall strategy adjustments.

[0074] In the healthcare business field, the intelligent agent can be an emergency vehicle, a mobile medical unit, or a telemedicine terminal. The matching decision-making tasks include patient transfer, on-site medical intervention, or remote diagnosis and treatment. During the execution process, medical resources are dispatched, patient data is collected, and medical operations are performed. The generated decision results include patient health status, medical intervention effects, and on-site data reports.

[0075] In the fintech business field, the intelligent agent can be an intelligent risk control system, an automated transaction control unit, or an account security module. The matching decision-making tasks include risk event isolation, account freezing, transaction interception, or system linkage operations. During the execution process, the agent analyzes account behavior in real time, identifies abnormal transactions, and performs security operations. The generated decision results include account status change records, risk disposal reports, and system feedback information.

[0076] In other different scenarios, the types and task forms of intelligent agents can be flexibly adjusted according to needs, making it suitable for complex and ever-changing environments such as manufacturing, urban management, and energy dispatch. The types of tasks performed cover physical operations, data processing, environmental control, and intelligent decision-making.

[0077] This embodiment processes the matching decision-making tasks performed by the intelligent agent. By combining the agent's real-time state with task requirements, it can efficiently and reliably complete diverse and dynamically complex tasks, ensuring that the system has the ability to respond quickly, execute stably, and provide intelligent feedback when facing emergencies, environmental changes, and multi-task collaboration. The standardized and structured output of task results further provides basic data for system optimization, model updates, and enhancement of agent capabilities, thereby improving the overall task execution efficiency, intelligent collaboration level, and dynamic adaptability of the system.

[0078] This invention relates to the field of artificial intelligence technology and can be applied to business scenarios such as fintech and healthcare. It discloses a method, apparatus, device, and medium for target event processing using language model collaboration, comprising: monitoring and identifying target events; collecting multi-source data associated with the target events; fusing the multi-source data to generate fused data; inputting the fused data into a pre-trained language model to generate a decision scheme; coordinating decisions based on the decision scheme; matching and assigning processing agents; utilizing the processing agents to execute the matched decision tasks; and generating decision results. This invention, by fusing multi-source data with deep inference using a pre-trained language model and combining it with a dynamic agent collaboration mechanism, achieves intelligent perception, rapid response, and collaborative processing of complex events, improving the system's adaptability and processing efficiency when dealing with unknown faults and complex tasks.

[0079] In one embodiment, step S10 above includes:

[0080] S101, Establish a full-dimensional monitoring matrix for the system's operating status, and obtain real-time operating status indicators through the full-dimensional monitoring matrix;

[0081] S102, identify abnormal behavior patterns in the real-time operating status indicators, and generate target event features based on the abnormal behavior patterns;

[0082] S103, perform event type classification and grading based on the target event characteristics, and generate classification and grading results including event category and severity level;

[0083] S104, Based on the classification and grading results, start the intelligent data acquisition engine, and collect associated multi-source data through the intelligent data acquisition engine;

[0084] S105, extract entity elements from the multi-source data, and mark the timestamp of the corresponding event for each entity element;

[0085] S106, Associate the entity element with the event type in the classification and grading results to generate type association data;

[0086] S107, Establish a relation topology based on the logical relationships between the entity elements;

[0087] S108, integrate the type-related data and relationship topology to generate an event data association graph containing timestamps.

[0088] In this embodiment, monitoring and identifying target events refers to capturing abnormal behaviors, sudden states, or potential risks in the system in real time through multi-dimensional information fusion, and accurately identifying them as specific events requiring further processing. This process relies on a multi-dimensional monitoring matrix of the system's operating status. This multi-dimensional monitoring matrix is ​​a collection of diverse information covering system hardware and software resources, network traffic, operation logs, user behavior, external threat information, and environmental factors. It possesses comprehensiveness and real-time capability, and can obtain real-time operating status indicators of the system through continuous monitoring, data aggregation, and dynamic analysis. Real-time operating status indicators include, but are not limited to, device performance parameters, network transmission indicators, system load information, resource utilization, and external security events. Data sources can include IoT devices, server terminals, application systems, sensor networks, and third-party data interfaces.

[0089] Based on real-time operational status indicators, statistical analysis, feature extraction, and pattern recognition methods are used to identify abnormal behavior patterns within these indicators. These abnormal behavior patterns include abnormal resource consumption, performance degradation, abnormal access behavior, system response latency, abnormal data fluctuations, and known threat patterns. Detection methods can employ a combination of rule-based matching, statistical thresholding, self-learning models, or anomaly detection algorithms to improve the accuracy and coverage of identification. Target event features are generated based on these abnormal behavior patterns. These features are structured information sets reflecting the nature, location, scope of impact, and evolution trend of abnormal states, including anomaly category, affected systems, detection time, preliminary assessment results, and historical correlation information.

[0090] Based on the characteristics of the generated target events, combined with classification criteria and risk models, the event type classification and grading operation is performed to form a classification and grading result that includes event category and severity level. The event category reflects the specific attributes and domain affiliation of the event, while the severity level describes the potential threat level of the event to system stability, business continuity or security. The grading criteria can be set according to industry standards, system requirements or dynamic adjustment strategies, and support flexible expansion and real-time adjustment.

[0091] Based on the classification and grading results, the intelligent data acquisition engine is activated. The intelligent data acquisition engine has the ability to be woken up on demand, dynamically configured, and acquire multi-source heterogeneous data. It can collect related multi-source data in a targeted manner according to the event category, scope of impact, and severity level. The sources of multi-source data include, but are not limited to, internal system logs, external environmental perception data, business operation information, user behavior records, and security monitoring results. The data types cover structured data, semi-structured data, and unstructured data, and support the integration of multiple communication protocols and data interfaces.

[0092] Entity elements are extracted from the collected multi-source data. Entity elements refer to key objects with independent attributes, clear identities, or traceability within the context of the target event, including device nodes, business modules, data resources, user accounts, and external entities. The extraction process combines data parsing, entity recognition, and attribute matching technologies to ensure the accuracy and completeness of entity elements. Each entity element is marked with a timestamp corresponding to the event that occurred. The timestamps use a standard time format to ensure consistency and comparability across systems and geographical locations.

[0093] By associating entity elements with event type information in the classification and grading results, type-related data is formed. This type-related data reflects the correspondence between entity elements and event attributes, event development stages, and risk levels, constructing a logical mapping from data entities to event semantics, thereby improving the efficiency of subsequent data analysis and reasoning.

[0094] Further, based on the logical relationships between entity elements, and combined with the interaction patterns, dependency structures, and communication links between entities, a relational topology is established. This relational topology, in the form of a graph structure, depicts the interactions and dependencies between entities, supporting dynamic updates and visual representation. By integrating type-related data with the relational topology, an event data association graph containing timestamps is formed. This event data association graph is characterized by its clear structure, intuitive relationships, and complete temporal information, providing a high-quality data foundation for subsequent deep analysis, pattern recognition, and intelligent reasoning.

[0095] This embodiment establishes a comprehensive monitoring matrix for the system's operational status, acquiring multi-dimensional real-time indicators and enhancing the perception range and monitoring accuracy of the system's status. Combined with abnormal behavior pattern recognition, it effectively improves the early detection capability of potential faults, system risks, and external threats. By generating target event features and classification results, it achieves standardized expression and risk level classification of complex events, facilitating subsequent resource scheduling and response strategy optimization. The intelligent data acquisition engine dynamically acquires multi-source data based on event characteristics, ensuring the relevance, completeness, and timeliness of data acquisition. Through entity element extraction and timestamp marking, a data traceability chain is constructed, enhancing the controllability and tracing capability of event analysis. The combination of type-related data and relational topology builds a structured and highly correlated data foundation. The resulting event data correlation graph possesses high availability and reliability, comprehensively improving the system's anomaly detection efficiency, data analysis depth, and complex event response capability.

[0096] In one embodiment, step S20 above includes:

[0097] S201, Perform data cleaning processing on the multi-source data to generate cleaned data;

[0098] S202, Perform format standardization conversion on the cleaned data to generate standardized data;

[0099] S203, extract feature vectors from the standardized data to generate a feature vector set;

[0100] S204, perform correlation mining based on the feature vector set to generate correlation data;

[0101] S205, perform data fusion processing on the aforementioned relationship data to generate fused data.

[0102] In this embodiment, multi-source data fusion processing refers to forming unified fused data that can be processed downstream by a series of structured, standardized, and correlation-enhancing operations based on data sets from multiple sources with different formats and complex structures. Multi-source data includes, but is not limited to, sensor data, log information, system monitoring data, user behavior data, external environment data, or other heterogeneous data sets with spatiotemporal attributes and business relevance. The sources can be internal information systems, network devices, third-party services, or other information collection channels, and the data types cover structured data, semi-structured data, and unstructured data.

[0103] First, multi-source data undergoes data cleaning. Data cleaning refers to removing redundant information, erroneous data, missing fields, inconsistent data formats, and outliers from the original data. This is achieved through methods such as missing data completion, error correction, anomaly detection, and noise filtering to improve the overall accuracy, completeness, and consistency of the data. The cleaning process can be combined with rule matching, pattern recognition, and statistical methods to verify and repair data from different sources according to predefined standards, ensuring that all types of data meet a unified quality standard. Data cleaning operations can be performed in batch processing mode or dynamically combined with real-time stream processing mechanisms.

[0104] Even after cleaning, data may still exhibit format differences and structural inconsistencies. Therefore, format standardization transformation must be performed on the cleaned data. Format standardization transformation refers to unifying data content from different sources, with different structures, or different encoding methods into a data expression form that conforms to preset data formats, data structures, and encoding standards. Standardization includes unifying field names, converting data types, correcting timestamp formats, standardizing units, and aligning encoding schemes to ensure efficient compatibility and accurate data interaction across different systems or modules. The standardization process can incorporate data model mapping, conversion rule bases, and dynamic parsing logic to adapt to various business needs and system environments.

[0105] Based on standardized data, feature vector extraction refers to the process of converting key attributes, implicit information, or multi-dimensional indicators in structured data into a set of statistically significant, expressive, and analytically valuable feature vectors using specific feature engineering methods. Feature vectors reflect the core information, important patterns, and potential relationships in the original data, and their types include numerical features, categorical features, time-series features, and combined features. The feature extraction process combines domain knowledge, data distribution, and business needs, and can apply statistical analysis, machine learning, or deep learning techniques to ensure that the generated feature vectors possess high discriminative power, high expressiveness, and good data adaptability.

[0106] Feature vector-based association mining refers to analyzing the inherent correlations, dependencies, or logical relationships between different data entities, data dimensions, or data fragments to form structured and semantically clear association data. The association mining process includes, but is not limited to, co-occurrence relationship identification, temporal dependency analysis, causal link reasoning, and pattern combination detection. It combines graph structure modeling, association rule analysis, or relationship network construction methods to reveal complex association patterns and potential structures between data.

[0107] Finally, the related data is fused. Data fusion refers to the unified integration, redundancy elimination, conflict coordination, and information supplementation of related data from different sources, with different structures, or different dimensions, to form fused data that is structurally complete, content-consistent, and clearly expressed. The fused data has the characteristics of high quality, high expressiveness, and strong correlation, and can be directly used for downstream analysis, reasoning, or model input, significantly improving the system's data utilization efficiency and intelligent processing capabilities.

[0108] Adaptive data cleaning strategies can be designed to address the differences between various data sources. For example, structural checks can be prioritized for log data, while outliers and missing values ​​can be addressed specifically for sensor data. The data cleaning module can dynamically configure a rule base, flexibly adjusting cleaning rules according to different business environments to improve cleaning efficiency and adaptability.

[0109] Standardized format conversion can combine a unified data exchange format (such as JSON, XML, Parquet) with data structure templates to ensure seamless data integration between different systems and interfaces. Feature vector extraction can employ statistical analysis methods, such as mean, variance, and distribution feature extraction, or it can be combined with machine learning models to extract complex high-order features, thereby improving feature representation capabilities.

[0110] In the process of relation mining, graph-based database-based graph structure storage and analysis solutions can be applied to achieve efficient processing of large-scale relation data. Relationship pattern recognition combined with rule mining and deep reasoning can adapt to the relation modeling needs in complex business scenarios. Data fusion processing can adopt a multi-stage fusion mechanism, first performing structural-level fusion, then semantic-level fusion and redundancy information optimization, to ensure the high reliability and business applicability of the final fused data.

[0111] This embodiment addresses the differences and conflicts in data quality, format, and information representation among multi-source heterogeneous data through a hierarchical, structured, and standardized fusion processing flow, thereby improving the overall consistency and usability of the data. After cleaning, standardization, and feature extraction, the multi-source data forms a high-quality, highly expressive set of feature vectors, significantly enhancing the effectiveness of data analysis and intelligent reasoning. Relationship mining further reveals the internal structure and potential patterns of the data. Data fusion processing eliminates data silos and redundant information, and the generated fused data provides accurate, complete, and high-value input for downstream intelligent models, ultimately improving the system's data processing efficiency, intelligent analysis capabilities, and business response speed.

[0112] In one embodiment, step S30 above includes:

[0113] S301, Perform integrity verification on the fused data and generate verified data;

[0114] S302, Map the verified data to a vector space to generate an input vector;

[0115] S303, Semantic relation parsing of the input vector is performed using a pre-trained language model to generate a semantic relation graph;

[0116] S304, extract contextual interaction patterns from the semantic relationship graph using the pre-trained language model to generate contextual features;

[0117] S305, Perform cross-domain analogy matching based on the context features using the pre-trained language model to generate an analogy reference set;

[0118] S306, construct a policy prototype based on the analog reference set using the pre-trained language model to generate policy prototype data;

[0119] S307, The pre-trained language model is used to optimize the constraints of the policy prototype data to generate a decision scheme.

[0120] In this embodiment, inputting fused data into the pre-trained language model means that, based on the data fusion processing results completed in the preceding steps, and relying on the trained language model structure and parameter system, the inference and generation of decision-making schemes are completed. The fused data, used as model input here, includes a structured, semantically related dataset after data cleaning, standardization, feature extraction, and relation mining, possessing completeness, consistency, and expressiveness.

[0121] First, the fused data undergoes integrity verification. Integrity verification, based on data quality management standards, ensures the input data is complete, formatted correctly, and semantically clear through structural checks, field validation, content consistency verification, and missing value detection. This prevents model inference errors caused by missing or abnormal data. Integrity verification can be implemented by combining rule engines, data model templates, and dynamic data monitoring mechanisms, covering multiple levels of verification at the field, record, and dataset levels. The output verified data meets the language model input requirements and possesses the characteristics of structural integrity, accurate content, and consistent encoding.

[0122] Mapping validated data to vector space refers to transforming structured and semantic data into multi-dimensional vector representations through operations such as feature encoding, numerical transformation, and high-dimensional mapping. The mapping process includes, but is not limited to, one-hot encoding of categorical data, normalization of numerical data, embedding representation of time-series data, and word vector modeling of text data. The final result is an input vector, which serves as the direct input to a pre-trained language model, reflecting the numerical, semantic, and structural features of the data.

[0123] The pre-trained language model performs semantic relation parsing on the input vector. Semantic relation parsing refers to extracting the semantic information, contextual associations, and potential logical relationships contained in the input vector based on the internal encoder structure of the language model, generating a structured semantic relation graph. The semantic relation graph expresses a multi-dimensional network of data entities, attributes, and relationships through a graph structure, reflecting the semantic dependencies, logical connections, and association paths between different data elements. The structure includes node information, edge weights, and multi-level relation labels, which facilitates subsequent in-depth analysis.

[0124] Based on semantic relationship graphs, pre-trained language models further extract contextual interaction patterns. Contextual interaction patterns refer to the dynamic semantic associations and logical interactions formed by input data at different dimensions and levels, including historical information, proximity relationships, and potential causal links. The extraction of contextual interaction patterns combines self-attention mechanisms, dependency structure analysis, and multi-layer semantic aggregation methods to generate contextual features. These features express the comprehensive semantic structure of the data at the temporal, spatial, and logical levels, providing fundamental information support for cross-domain reasoning and complex analysis.

[0125] By using a pre-trained language model, cross-domain analogy matching is performed based on contextual features. Cross-domain analogy matching refers to aligning and semantically mapping the semantic features of the current input data with the knowledge system, historical case library, or multi-domain data samples stored within the model to identify similar structures, patterns, and potential logical connections, forming an analogy reference set. The analogy reference set contains cases, patterns, and knowledge structures that are highly similar to the current input, improving the system's adaptability to unknown scenarios, novel problems, and complex tasks.

[0126] The pre-trained language model constructs a strategy prototype based on an analog reference set. The strategy prototype is a preliminary solution structure with feasibility, executability and optimization space, formed based on historical experience, knowledge structure and analogical reasoning results. The strategy prototype data includes strategy objectives, task paths, resource allocation and constraints.

[0127] Finally, the pre-trained language model is used to optimize the constraints of the strategy prototype data. Constraint optimization refers to adjusting, improving and optimizing the execution path, resource configuration and parameter settings in the strategy prototype by combining the current system state, resource capabilities and external environment, so as to generate a complete, optimized and implementable decision solution. The decision solution is targeted, flexible and highly adaptable, and meets the actual business needs and system operating conditions.

[0128] This embodiment ensures the accuracy, completeness, and consistency of input data content through multi-level, structured data verification, vector mapping, and semantic parsing, reducing error rates and uncertainties in the model inference process. The construction of semantic relationship graphs and contextual interaction patterns enhances the semantic expressiveness and multi-dimensional logical connections of the data, improving the system's ability to understand complex, dynamic, and cross-domain problems. A cross-domain analogy matching mechanism enables the system to effectively identify historical patterns and knowledge structures similar to unknown scenarios, enhancing the reasoning depth and generalization ability of the decision-making process. The construction of strategy prototypes and the optimization of constraints ensure that the final generated decision scheme is targeted, executable, and dynamically adaptable, comprehensively improving the system's intelligent response efficiency and processing accuracy for complex events, unexpected problems, and cross-domain tasks.

[0129] In one embodiment, step S40 above includes:

[0130] S401, parse the task elements of the decision scheme to generate a decision task set containing all tasks to be assigned;

[0131] S402, Analyze the execution dependencies of each decision task in the decision task set and generate a task dependency topology graph;

[0132] S403, Analyze the real-time capability status of currently available processing agents and generate a capability status report that includes load and capability indicators;

[0133] S404, Based on the task dependency topology graph and the capability status report, perform dynamic task matching processing to generate a matching result between the decision task and the agent;

[0134] S405, optimize the resource allocation strategy based on the matching result between the decision task and the agent, and generate the final agent allocation scheme.

[0135] In this embodiment, decision coordination is performed based on a decision scheme, aiming to achieve dynamic resource allocation and multi-agent collaboration in complex task environments. First, the task elements of the decision scheme are analyzed. These elements include the task objectives, execution conditions, resource requirements, and time constraints described in the decision scheme. The analysis process can be based on text parsing, structured mapping, and semantic analysis methods to transform unstructured information in the decision scheme into structured task data, ensuring the complete expression and accurate extraction of task information. The specific content of the task elements includes the task name, functional description, dependencies, input-output relationships, time requirements, and resource types, providing a clear data foundation for subsequent task organization and allocation.

[0136] By systematically organizing task elements, a decision task set containing all tasks to be assigned is generated. The decision task set is a structured task list that comprehensively covers all types of decision tasks that need to be executed. Each task in the set has a unique identifier, clear execution conditions and dependency information, which facilitates the organization, coordination and dynamic adjustment of the overall tasks.

[0137] Further analysis reveals the execution dependencies among the decision tasks in the decision task set. These dependencies reflect the logical connections and sequential constraints between tasks in terms of data flow, function calls, and execution order. The dependency analysis is based on the task data structure, business logic, and operation chain, combined with task input-output relationships, time constraints, and functional dependencies, to generate a task dependency topology graph. This graph uses a graph structure to represent task nodes and dependent edges, clearly depicting the overall task structure, concurrency relationships, and execution paths, facilitating task decomposition, path optimization, and conflict detection.

[0138] After task organization is completed, the real-time capability status of currently available processing agents is analyzed. This capability status reflects the real-time load, resource consumption, performance parameters, and health status of each processing agent. The analysis process obtains multi-dimensional status information through a status monitoring module, resource management system, and internal agent feedback mechanisms. The generated capability status report includes the load level, available resources, functional matching degree, and historical performance of each processing agent. Capability metrics are set in conjunction with multi-dimensional parameters such as CPU utilization, memory usage, task response time, error rate, and task completion rate to ensure the accuracy and real-time nature of agent allocation decisions.

[0139] Dynamic task matching is performed based on task dependency topology graphs and capability status reports. This dynamic matching combines task requirements, execution dependencies, and agent capabilities. Through optimization algorithms, matching rules, and real-time evaluation, the optimal task-agent mapping relationship is determined, generating decision-making task-agent matching results. The matching results clearly define the optimal executing agent for each task, considering resource utilization, load balancing, execution efficiency, and system stability. The matching scheme is dynamically adjusted to adapt to task changes and resource status fluctuations.

[0140] Based on the matching results, the resource allocation strategy is optimized. This strategy is a collaborative optimization plan addressing the overall system's resource and task requirements, encompassing specific agent allocation, resource reservation, task concurrency adjustment, and load balancing strategies. The final agent allocation plan is a structured output based on the overall system perspective, dynamic information, and optimization results. The plan clearly defines the executing entity, resource allocation, timing, and reservation mechanisms for each task, ensuring efficient task execution, rational resource utilization, and stable system operation.

[0141] This embodiment ensures complete and clear multi-task information in complex decision-making schemes through in-depth analysis and structured expression of task elements, improving the efficiency of task organization and management. The construction of a task dependency topology graph intuitively displays the task structure and logical relationships, optimizing task execution paths and reducing task conflicts and dependency errors. The generation of capability status reports provides real-time and comprehensive agent capability information, supporting accurate and efficient task allocation decisions. The dynamic task matching mechanism, based on real-time data and optimization algorithms, flexibly adapts to changes in tasks and resources, improving resource utilization and system response speed. System optimization of resource allocation strategies ensures efficient overall resource allocation, load balancing, and execution stability, enhancing the system's adaptability and collaborative processing capabilities in complex and changing environments.

[0142] In one embodiment, step S50 above includes:

[0143] S501, Distribute task execution instructions for corresponding decision-making tasks to the corresponding processing agents according to the agent allocation scheme, and generate a task instruction set;

[0144] S502, the processing agent executes the operation commands in the task instruction set to generate an operation execution record;

[0145] S503, Monitor the task execution progress and resource status of the processing agent, and generate a status monitoring dataset;

[0146] S504, based on the status monitoring dataset, detect abnormal execution behavior and generate an anomaly detection report;

[0147] S505, aggregate the operation execution records, status monitoring dataset, and anomaly detection report to generate a decision result that includes quantitative analysis of the execution effect.

[0148] In this embodiment, decision results are generated by processing agents executing matched decision tasks. Specifically, this includes distributing task execution instructions for the corresponding decision tasks to the corresponding processing agents based on an agent allocation scheme. The agent allocation scheme originates from the preceding task matching process, clarifying the execution entity, required resources, and time requirements for each task, ensuring that the mapping relationship between tasks and agents conforms to the current system capacity and load. The process of distributing task execution instructions is based on a distributed communication mechanism or a local instruction scheduling system, encapsulating task content, execution parameters, and associated data into a structured task instruction set to ensure information integrity and transmission reliability. The task instruction set includes a unique task identifier, functional description, input data, operation sequence, and expected output results, facilitating accurate reception and efficient execution by the processing agents.

[0149] By processing the operation commands in the task instruction set executed by the intelligent agent, an operation execution record is generated. The operation commands are the specific execution content within the task instruction set, encompassing various operation types such as function calls, logical operations, data processing, and environmental interactions. The processing intelligent agent sequentially parses and executes the operation commands according to the instruction set, forming complete data of the task execution process. The operation execution record details the execution time, operation steps, resource consumption, return results, and exception information, possessing real-time, traceability, and structured characteristics, facilitating subsequent analysis and system management.

[0150] During task execution, the task progress and resource status of the processing agent are continuously monitored, generating a status monitoring dataset. Task progress includes task start time, completion percentage, interim results, and overall execution time. Resource status encompasses metrics such as CPU utilization, memory usage, network latency, storage space, and feedback from the agent's internal sensors. The status monitoring dataset is a multi-dimensional, real-time, dynamic data set reflecting the real-time status and operational quality of the processing agent during task execution, supporting the system in dynamically adjusting task strategies based on status changes.

[0151] Based on the status monitoring dataset, abnormal behavior detection is performed, generating anomaly detection reports. Anomaly detection identifies abnormal situations during task execution, such as timeouts, resource anomalies, error returns, functional failures, or security threats, through rule matching, model analysis, and real-time comparison. The anomaly detection report includes the anomaly type, time, location, scope of impact, and preliminary cause analysis, providing accurate data support for subsequent troubleshooting and system optimization.

[0152] Finally, the system aggregates operation execution records, status monitoring datasets, and anomaly detection reports to generate decision results that include quantitative analysis of execution effectiveness. This quantitative analysis, based on preset evaluation indicators, comprehensively considers task completion, resource consumption, response time, accuracy, and stability to form a multi-dimensional performance evaluation. The decision results are output in structured data format, including execution status, performance indicators, anomaly information, and performance scores, providing reliable data for closed-loop optimization, model adjustment, and agent policy updates.

[0153] This embodiment ensures that task content and parameters are accurately transmitted to the corresponding processing agent through a task execution instruction distribution mechanism, improving execution accuracy and response efficiency. Systematic generation of operation execution records enables full-process data retention, enhancing task traceability and auditability. Real-time status monitoring and multi-dimensional data acquisition provide a comprehensive understanding of the processing agent's operational status, promptly identifying performance bottlenecks and resource anomalies to ensure stable system operation. An abnormal behavior detection mechanism proactively identifies potential faults and security risks, providing early warnings and supporting rapid response, reducing system failure rates and security vulnerabilities. Quantitative analysis of execution results uses multi-dimensional indicators to comprehensively evaluate task execution results, providing objective and data-driven feedback, promoting system self-optimization and dynamic adjustment, and improving overall processing efficiency, resource utilization, and system stability.

[0154] In one embodiment, after step S50 above, the method further includes:

[0155] S601, Extract quantitative features of execution effect from the decision results;

[0156] S602, Based on the quantitative characteristics of the execution effect, analyze the success and failure modes of the decision-making process and generate a decision-making mode analysis report;

[0157] S603, Perform spatiotemporal correlation analysis on the decision-making pattern analysis report and the fused data to generate spatiotemporally correlated fused correlation data;

[0158] S604, Based on the fused and correlated data, generate language model parameter update data and process agent behavior strategy update data;

[0159] S605, update the inference parameters of the pre-trained language model using the language model parameter update data, and update the decision policy library of the processing agent using the processing agent behavior policy update data.

[0160] In this embodiment, after the intelligent agent executes a matching decision task and generates a decision result, quantitative features of the execution effect are extracted from the decision result. The decision result comes from the operation execution record, status monitoring dataset, and anomaly detection report during the task execution process. The process of extracting quantitative features of the execution effect includes statistical calculation, feature modeling, and index transformation of the above data. Quantitative features of the execution effect include, but are not limited to, task completion rate, response time, resource consumption, error rate, system stability, fault recovery time, and anomaly occurrence frequency. The specific extraction process uses aggregation calculation, normalization processing, and multi-dimensional feature combination to ensure that the feature data has accuracy, quantifiability, and a structured expression suitable for subsequent analysis.

[0161] Based on the quantitative characteristics of execution effectiveness, this study analyzes success and failure patterns in decision-making, generating a decision pattern analysis report. The analysis process utilizes statistical modeling, pattern recognition, and historical data comparison to uncover differences in the effectiveness, performance trends, and failure patterns of different decision-making tasks under varying conditions. Successful decision-making patterns include efficient task completion, rational resource utilization, and stable system operation; failure patterns include execution anomalies, resource bottlenecks, performance degradation, or system outages. The decision pattern analysis report presents structured data and visual charts, including typical patterns, influencing factors, related parameters, and improvement suggestions to support subsequent system optimization and strategy adjustments.

[0162] The decision-making pattern analysis report and fused data are subjected to spatiotemporal correlation analysis to generate spatiotemporally correlated fused data. The fused data originates from the preceding multi-source data fusion process and includes time information, spatial location, environmental state, and system background information related to the target event. Based on time synchronization, spatial matching, and logical relationship reasoning, the spatiotemporal correlation analysis establishes a dynamic relationship between the decision-making pattern and system state and environmental changes, generating spatiotemporally correlated fused data that reflects the system's operational patterns and changes in decision-making effects under different time and spatial conditions, thereby improving the overall system's environmental adaptability and dynamic analysis capabilities.

[0163] Based on fused and correlated data, language model parameter update data and agent behavior policy update data are generated. The language model parameter update data optimizes model parameters related to reasoning ability, contextual understanding, and output accuracy in the pre-trained language model, supporting fine-tuning of model structure, weight adjustment, and supplementation of training data. The agent behavior policy update data provides new behavioral rules, parameter configurations, and policy optimization paths for task execution logic, resource allocation strategies, and anomaly response mechanisms. The update data generation process is based on data analysis, performance evaluation, and optimization algorithms to ensure the scientific validity, practicality, and system compatibility of the updated content.

[0164] Update data using language model parameters to improve the inference parameters of the pre-trained language model, thereby enhancing the model's accuracy and adaptability in multi-source data understanding, complex reasoning, and dynamic decision-making. Update data using the processing agent's behavioral policies to update the agent's decision policy library, optimizing the agent's decision logic and execution efficiency in task execution, resource scheduling, and anomaly handling, and enhancing the system's overall self-learning, adaptive, and continuous optimization capabilities.

[0165] Example Description: In a hospital intelligent operation and maintenance management system, the first step is to conduct comprehensive status monitoring of the hospital information system, medical equipment, environmental monitoring system, and network transmission platform. This involves monitoring and collecting real-time operational status indicators from the electronic medical record system, drug management system, surgical scheduling system, network switching equipment, vital sign monitors, medical imaging equipment, and building control system. Key data such as processing latency, equipment load, data transmission delay, equipment alarm logs, and changes in environmental temperature and humidity are continuously extracted. Anomaly detection algorithms identify issues such as abnormal fluctuations in ECG monitoring data, abnormal data transmission in the automated drug dispensing system, abnormal latency in the medical imaging transmission network, and excessive ward temperatures. Corresponding abnormal data features are extracted to form structured feature information for the target events. For different abnormal behaviors, event type identification and grading algorithms are used to classify equipment anomalies, environmental anomalies, and system latency anomalies, determining different event categories such as equipment failure, environmental safety, or information transmission risks. Severity levels are then assigned based on the scope and urgency of the anomaly's impact. Based on the categorized event types and severity levels, the intelligent data acquisition engine in the medical environment is automatically activated to synchronously collect historical data, real-time data, and upstream and downstream interaction data before and after the occurrence of abnormal events from IoT devices, information system logs, device control records, and maintenance work order databases. The collected multi-source data is timestamped, recording the time of device status change, the time of abnormal event triggering, and the time of data transmission delay. Device type, system node, data packet flow, and abnormal category in entity elements are associated to generate type-related data between devices and event categories. Simultaneously, based on the dependency paths between devices and the system architecture, an operational relationship topology between devices is established. Integrating the type-related data and relationship topology, a hospital information system abnormal event data association graph containing time stamps and device logical paths is generated.

[0166] During data fusion, data cleaning is performed on the collected medical device data, network logs, environmental monitoring data, and system operation records to remove idle data, irrelevant logs, and duplicate records, retaining data highly relevant to the target events. Through format standardization, data from different systems and devices are uniformly converted into structured table formats, including standardized device status fields, environmental parameter fields, and log operation sequences, ensuring a unified parsing format across the data. Key feature vectors of medical device operating status are extracted, including device power consumption levels, data processing latency, task queue length, number of abnormal alarms, and data transmission rate, forming a feature vector set. Multi-dimensional correlation analysis is performed based on this feature vector set to identify potential correlations between medical devices and environmental changes, and dependency paths between device operation and network load, generating correlation data of device operating status. This correlation data is then fused, eliminating redundant paths, abnormal noise, and unstable connections, ultimately generating complete and time-sequentially ordered fused data of medical device operating status.

[0167] The fused data is input into a pre-trained language model specifically designed for intelligent hospital operation and maintenance. First, the fused data undergoes integrity verification, detecting missing equipment data, abnormal time tags, or log fragments with structural anomalies. The verified data is then mapped into a vector space to generate input vectors suitable for language model inference. The pre-trained language model is used to perform semantic relationship parsing of the medical equipment's operating status on the input vectors, generating a semantic relationship graph between equipment status, system logs, environmental parameters, and task scheduling data. Further, contextual interaction patterns arising from equipment status changes, data transmission path switching, and environmental fluctuations are extracted from the semantic relationship graph, extracting the interaction features between the equipment, system, and environment. The language model performs cross-device analogy inference on these interaction features, retrieving similar historical anomaly handling cases and comparing them with the current equipment anomaly scenario to generate an analogy reference set. Based on this analogy reference set, the language model constructs prototype medical equipment anomaly handling strategies, including priority path switching strategies for data latency, task migration strategies for excessive equipment load, and mandatory alarm strategies for exceeding environmental limits, forming prototype strategy data for optimization. Through constraint optimization, combined with the current hospital equipment status and task priorities, a decision-making scheme suitable for the current scenario is generated.

[0168] The generated decision-making scheme is analyzed, and all pending medical device task elements are extracted to form a decision task set, including different types of processing tasks such as device restart, task migration, path switching, and anomaly reporting. The execution dependencies between tasks in the task set are analyzed, a task dependency topology graph is established, and the dependency order of device restart (which must be performed after path switching) and task migration (which must be performed after device load adjustment) is clarified. The current available processing agent capabilities are analyzed in real time, including the task processing load of the hospital information management system, the instruction execution capability of the device control module, and the data path switching speed, forming current agent load and capability indicators. Based on the task dependency topology graph and agent capability status, dynamic task matching is performed, prioritizing the allocation of device restart tasks to idle processing modules and path switching tasks to network path control agents, generating the current task-agent matching results. Based on the matching results, resource allocation strategies are optimized, processing order is adjusted, resource conflicts are avoided, and the final agent allocation scheme is generated.

[0169] According to the agent allocation scheme, task execution instructions for target devices are sent to different processing agents, forming a task instruction set, which includes device restart instructions, path switching instructions, task migration instructions, and anomaly alarm instructions. The processing agents execute the operation commands in the task instruction set, recording device status changes, task migration progress, path switching success rate, and anomaly notification response time, generating detailed operation execution records. The progress and resource consumption of the agents during task execution are continuously monitored, collecting task completion time, data transmission load, and processing module utilization, generating a status monitoring dataset. Abnormal behaviors during task execution, such as path switching failures, task migration delays, and device restart anomalies, are detected in real time, generating anomaly detection reports. Finally, the operation execution records, status monitoring dataset, and anomaly detection reports are aggregated to generate decision results that include quantitative analysis of execution effectiveness.

[0170] Quantitative features of execution effectiveness are extracted from the decision-making results, focusing on task completion rate, path switching success rate, task migration response time, equipment restart and recovery time, and anomaly notification processing latency. Based on the extracted quantitative features, success and failure modes of task processing are analyzed, generating a decision-making pattern analysis report that includes equipment type, task type, operation path, and success probability. The decision-making pattern analysis report and fused data are subjected to spatiotemporal correlation analysis to identify differences in task processing effectiveness under different equipment operating states, environmental parameters, and network loads, generating spatiotemporally correlated fused correlation data. Language model parameter update data is generated based on the correlation data, adjusting the model's inference weights for specific equipment states and the priority of anomaly handling paths, generating agent behavior strategy update data, and optimizing equipment task scheduling logic and anomaly response rules. The language model parameter update data is used to update the inference parameters of the pre-trained language model, and the agent behavior strategy update data is used to update the decision-making strategy library of the agent, improving the system's adaptability, response efficiency, and processing accuracy in subsequent handling of hospital equipment anomalies and network congestion events.

[0171] In the bank's intelligent risk management system, comprehensive operational status monitoring is first implemented for the credit business system, transaction monitoring system, payment and settlement system, and customer service platform. A multi-dimensional monitoring matrix of financial business operation status is constructed by collecting data on transaction delays, system load, anomaly alarms, interface response times, customer complaint records, and real-time fund flow status. Real-time anomaly detection algorithms identify behavioral patterns such as abnormal transaction peak delays, fund path blockages, high-frequency payment failures, and core system response timeouts. Key feature information such as transaction numbers, fund flows, account identifiers, and system loads from these anomalies is extracted to form a structured description of the target event. For different types of financial anomalies, an event type identification and grading mechanism distinguishes different event categories such as transaction timeouts, payment failures, data synchronization anomalies, and account risks. The severity level of the event is then classified based on the scope of the anomaly's impact, the degree of fund risk exposure, and the urgency of the business. Based on the classification results, the financial data acquisition engine is automatically activated to synchronously collect historical transactions, real-time fund flows, anomaly alarms, and interface call information from the transaction log database, risk control logs, account anomaly records, payment channel status information, and third-party interface logs. The collected multi-source data is timestamped to record the time of abnormal transaction generation, the time of payment interface anomaly occurrence, and the response time of fund freezing operation. Transaction records, account status, and anomaly types are correlated to generate type-related data. At the same time, a logical relationship topology is established based on fund paths, account relationships, and transaction chains. The type-related data and relationship topology are integrated to form a data correlation graph of financial transaction anomaly events that includes timestamp markers and path relationships.

[0172] During data fusion, the collected financial transaction data, risk monitoring logs, account operation records, and payment system interface logs undergo data cleaning to remove redundant transactions, irrelevant operations, and invalid states, retaining only data directly related to the target anomaly. Through format standardization, transaction logs, interface logs, and account status information from different business systems are uniformly converted into a structured data format, ensuring consistent field definitions for transaction time, account information, fund flow, and anomaly type. Key feature vectors affecting transaction efficiency and account security are extracted, including transaction processing delays, anomaly alarm frequency, fund path congestion, and payment failure rate, forming a feature vector set. Based on this feature vector set, correlation mining is performed to identify fund flow paths, account risk propagation paths, and high-load nodes in the system, generating correlation data for the financial system's anomaly handling chain. This correlation data is then fused, removing redundant anomaly paths and unstable transaction nodes to generate a coherent and stable financial business data fusion result.

[0173] The fused data is input into a pre-trained language model for financial intelligent risk control. First, the fused data undergoes integrity verification to detect missing transaction data, abnormal account information, or incomplete path links. The verified data is then mapped into a vector space to generate input vectors for semantic processing in the financial domain. The pre-trained language model parses the semantic relationships of the input vectors along transaction links, constructing a semantic relationship graph of account relationships, transaction paths, payment nodes, and risk propagation paths. Further, contextual interaction patterns of transaction delays, fund freezes, and payment anomalies are extracted from the semantic relationship graph, generating interaction features between multi-account fund flows and transaction failures. Based on these interaction features, the pre-trained language model performs cross-account and cross-payment system analogical reasoning, retrieving similar historical risk events and current transaction anomaly scenarios to generate an analogy reference set. Based on this analogy reference set, strategy prototypes are constructed, including high-load path switching strategies, transaction priority adjustment strategies, and automatic compensation strategies for abnormal payments, forming strategy prototype data. Finally, constraint optimization techniques are applied, comprehensively considering the current account risk level, transaction path load, and payment node response status to generate a currently applicable risk management decision-making scheme.

[0174] The system analyzes decision-making schemes, extracting risk control tasks such as account freezing, payment channel switching, transaction rate limiting, and high-risk transaction marking, forming a decision task set. It analyzes the task execution dependencies within this set, generating a task dependency topology graph to clarify the execution order: account freezing must take precedence over payment channel switching, and transaction rate limiting must be executed only after account risk confirmation. The system analyzes the real-time capability status of available agents, assessing the current load and processing capacity of the payment system's channel control module, transaction processing module, and risk isolation module, generating a capability status report based on load and capability indicators. Dynamic task matching is performed based on the task dependency topology graph and capability status report, prioritizing the allocation of high-risk account freezing tasks to idle risk control agents and payment channel switching tasks to high-bandwidth channel control agents, generating task-agent matching results. Based on the matching results, the system optimizes resource allocation strategies, balancing the load of each agent to avoid task congestion, generating the final agent allocation scheme.

[0175] According to the agent allocation scheme, task execution instructions are distributed to different processing agents, generating a task instruction set that includes instructions such as account freezing, payment channel switching, transaction rate limiting, and high-risk transaction marking. The processing agents execute the commands in the task instruction set, recording task execution time, account status changes, payment channel switching results, and risk isolation success rates, generating an operation execution log. The progress and resource status of the processing agents during task execution are continuously monitored, collecting data on task completion latency, system resource consumption, and payment channel load changes, generating a status monitoring dataset. Abnormal behaviors during task execution are detected in real time, such as failed account freezing, abnormal payment path switching, and missed high-risk transaction marking, generating an anomaly detection report. Finally, the operation execution log, status monitoring dataset, and anomaly detection report are aggregated to generate a decision result that includes quantitative analysis of the execution effect.

[0176] Quantitative features of execution performance, such as task completion rate, risk control response time, payment path recovery time, and accuracy of abnormal transaction handling, are extracted from the decision-making results. Based on these extracted features, success and failure patterns in task processing are analyzed, generating a decision-making pattern analysis report that includes account type, transaction path, task type, and success probability. The decision-making pattern analysis report and fused data are then subjected to spatiotemporal correlation analysis to identify differences in task processing performance under different account structures, transaction paths, and payment node loads, generating spatiotemporally correlated financial transaction data fusion results. Based on the fusion results, language model parameter update data is generated to adjust the model's inference weights for specific account risks, payment path anomalies, and transaction delays. Simultaneously, processing agent behavior strategy update data is generated to optimize execution strategies for account freezing, pathway switching, and transaction rate limiting. The language model parameter update data is used to update the inference parameters of the pre-trained language model, and the processing agent behavior strategy update data is used to update the decision-making strategy library of the processing agent, improving the real-time adaptability and risk handling efficiency of the financial business system to new transaction risks, account anomalies, and payment congestion.

[0177] This embodiment extracts quantitative features of execution effectiveness from decision-making results, forming a measurable and traceable performance data foundation, thereby enhancing the system's ability to objectively evaluate decision quality and execution effectiveness. Based on quantitative feature analysis of decision success and failure patterns, the system can accurately identify efficient solutions and potential risks, promoting experience accumulation and decision optimization. Spatiotemporal correlation analysis integrates system state, environmental factors, and decision effects, comprehensively reflecting the system's dynamic operating characteristics and improving environmental adaptability and decision accuracy. Generating language model parameter update data and processing agent behavior strategy update data enables dynamic collaborative optimization of model inference capabilities and agent execution strategies. The system possesses continuous learning, real-time optimization, and self-adaptation capabilities, reducing the need for manual intervention, improving overall system processing efficiency, decision-making level, and stability, and enhancing its ability to cope with complex, dynamic, and unknown scenarios.

[0178] In one embodiment, a target event processing apparatus for language model collaboration is provided, which corresponds one-to-one with the target event processing method for language model collaboration described in the above embodiments. (Refer to...) Figure 3 , Figure 3 This is a schematic diagram of the functional modules of a preferred embodiment of the language model collaborative target event processing device of the present invention. The modules include an event monitoring module 10, a data fusion module 20, a language model inference module 30, an agent coordination module 40, and a task execution module 50. Detailed descriptions of each functional module are as follows:

[0179] Event monitoring module 10 is used to monitor and identify target events and collect multi-source data associated with the target events;

[0180] Data fusion module 20 is used to fuse the multi-source data to generate fused data;

[0181] The language model inference module 30 is used to input the fused data into a pre-trained language model and generate a decision scheme through the pre-trained language model.

[0182] The agent coordination module 40 is used to coordinate decisions according to the decision scheme, and to match and assign processing agents for the decision tasks in the decision scheme.

[0183] The task execution module 50 is used to execute the matched decision task through the processing agent and generate decision results.

[0184] In one embodiment, the event monitoring module 10 is specifically used for:

[0185] Establish a full-dimensional monitoring matrix for system operation status, and obtain real-time operation status indicators through the full-dimensional monitoring matrix;

[0186] Identify abnormal behavior patterns in the real-time operating status indicators, and generate target event features based on the abnormal behavior patterns;

[0187] Based on the characteristics of the target event, perform event type classification and grading to generate classification and grading results that include event category and severity level;

[0188] Based on the classification and grading results, the intelligent data acquisition engine is activated, and associated multi-source data is collected through the intelligent data acquisition engine.

[0189] Extract entity elements from the multi-source data and mark each entity element with the timestamp of the corresponding event.

[0190] The entity elements are associated with the event types in the classification and grading results to generate type association data;

[0191] Establish a relation topology based on the logical relationships between the entity elements;

[0192] By integrating the aforementioned types of associated data and relationship topology, an event data association graph containing timestamps is generated.

[0193] In one embodiment, the data fusion module 20 is specifically used for:

[0194] The multi-source data is cleaned to generate cleaned data;

[0195] The cleaned data is then subjected to format standardization conversion to generate standardized data;

[0196] Feature vectors are extracted from the standardized data to generate a feature vector set;

[0197] Based on the feature vector set, correlation mining is performed to generate correlation data;

[0198] The data relationship data is fused to generate fused data.

[0199] In one embodiment, the language model inference module 30 is specifically used for:

[0200] The fused data is subjected to integrity verification to generate verified data;

[0201] The verified data is mapped to a vector space to generate an input vector.

[0202] The semantic relationships of the input vector are parsed using a pre-trained language model to generate a semantic relationship graph.

[0203] The pre-trained language model extracts contextual interaction patterns from the semantic relationship graph to generate contextual features;

[0204] The pre-trained language model performs cross-domain analogy matching based on the context features to generate an analogy reference set.

[0205] The policy prototype data is generated by constructing a policy prototype based on the analog reference set using the pre-trained language model.

[0206] The pre-trained language model is used to optimize the constraints of the policy prototype data to generate a decision scheme.

[0207] In one embodiment, the agent coordination module 40 is specifically used for:

[0208] Analyze the task elements of the decision-making scheme to generate a decision task set containing all tasks to be assigned;

[0209] Analyze the execution dependencies of each decision task in the decision task set and generate a task dependency topology graph;

[0210] Analyze the real-time capability status of currently available processing agents and generate a capability status report that includes load and capability metrics;

[0211] Based on the task dependency topology graph and the capability status report, dynamic task matching processing is performed to generate a matching result between the decision task and the agent.

[0212] Based on the matching results between the decision-making task and the agent, optimize the resource allocation strategy and generate the final agent allocation scheme.

[0213] In one embodiment, the task execution module 50 is specifically used for:

[0214] Based on the agent allocation scheme, task execution instructions for the corresponding decision-making agents are distributed to the corresponding processing agents to generate task instruction sets;

[0215] The processing agent executes the operation commands in the task instruction set to generate an operation execution record.

[0216] Monitor the task execution progress and resource status of the processing agent, and generate a status monitoring dataset;

[0217] Based on the aforementioned status monitoring dataset, abnormal execution behaviors are detected, and an anomaly detection report is generated.

[0218] By aggregating the operation execution records, status monitoring datasets, and anomaly detection reports, a decision result containing quantitative analysis of the execution effect is generated.

[0219] In one embodiment, the task execution module 50 is specifically used for:

[0220] Extract quantitative features of execution effectiveness from the decision results;

[0221] Based on the quantitative characteristics of the execution effect, analyze the success and failure patterns of decision-making and generate a decision-making pattern analysis report;

[0222] The decision-making pattern analysis report and the fused data are subjected to spatiotemporal correlation analysis to generate spatiotemporally correlated fused correlation data;

[0223] Based on the fused and correlated data, language model parameter update data and intelligent agent behavior strategy update data are generated;

[0224] The inference parameters of the pre-trained language model are updated using the language model parameter update data, and the decision policy library of the processing agent is updated using the processing agent behavior policy update data.

[0225] In one embodiment, a computer device is provided, which may be a server, and its internal structure diagram may be as follows: Figure 4 As shown. The computer device includes a processor, memory, network interface, and database connected via a system bus. The processor provides determination and control capabilities. The memory includes non-volatile and / or volatile storage media and internal memory. The non-volatile storage media stores the operating system, computer programs, and database. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage media. The network interface is used for communication with external user terminals via a network connection. When the computer program is executed by the processor, it implements the server-side functions or steps of a language model-coordinated target event processing method.

[0226] In one embodiment, a computer device is provided, which may be a user terminal, and its internal structure diagram may be as follows: Figure 5 As shown, the computer device includes a processor, memory, network interface, display screen, and input devices connected via a system bus. The processor provides determination and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage media. The network interface is used to communicate with an external server via a network connection. When the computer program is executed by the processor, it implements the user-side functions or steps of a language model-coordinated target event handling method.

[0227] In one embodiment, a computer device is provided, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to perform the following steps:

[0228] Monitor and identify target events, and collect multi-source data associated with the target events;

[0229] The multi-source data is fused to generate fused data;

[0230] The fused data is input into a pre-trained language model, and a decision scheme is generated through the pre-trained language model.

[0231] Decision coordination is performed according to the decision scheme, and processing agents are matched and assigned to the decision tasks in the decision scheme.

[0232] The processing agent executes the matching decision-making task and generates the decision result.

[0233] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon, the computer program performing the following steps when executed by a processor:

[0234] Monitor and identify target events, and collect multi-source data associated with the target events;

[0235] The multi-source data is fused to generate fused data;

[0236] The fused data is input into a pre-trained language model, and a decision scheme is generated through the pre-trained language model.

[0237] Decision coordination is performed according to the decision scheme, and processing agents are matched and assigned to the decision tasks in the decision scheme.

[0238] The processing agent executes the matching decision-making task and generates the decision result.

[0239] It should be noted that the functions or steps that can be implemented by the computer-readable storage medium or computer device described above can be referred to the relevant descriptions on the server side and user side in the foregoing method embodiments. To avoid repetition, they will not be described one by one here.

[0240] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. Any references to memory, storage, databases, or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), Rambus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0241] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the above-described division of functional units and modules is used as an example. In practical applications, the above functions can be assigned to different functional units and modules as needed, that is, the internal structure of the device can be divided into different functional units or modules to complete all or part of the functions described above.

[0242] It should be noted that if any software tools or components not belonging to this company appear in the embodiments of this application, they are merely illustrative examples and do not represent actual use. The embodiments described above are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should all be included within the protection scope of the present invention.

Claims

1. A method for handling target events in a language model collaboration framework, characterized in that, Includes the following steps: Monitor and identify target events, and collect multi-source data associated with the target events; The multi-source data is fused to generate fused data; The fused data is input into a pre-trained language model, and a decision scheme is generated through the pre-trained language model. Decision coordination is performed according to the decision scheme, and processing agents are matched and assigned to the decision tasks in the decision scheme. The processing agent executes the matching decision-making task and generates decision results.

2. The target event processing method for language model collaboration as described in claim 1, characterized in that, Monitor and identify target events, and collect multi-source data associated with the target events, including: Establish a full-dimensional monitoring matrix for system operation status, and obtain real-time operation status indicators through the full-dimensional monitoring matrix; Identify abnormal behavior patterns in the real-time operating status indicators, and generate target event features based on the abnormal behavior patterns; Based on the characteristics of the target event, perform event type classification and grading to generate classification and grading results that include event category and severity level; Based on the classification and grading results, the intelligent data acquisition engine is activated, and associated multi-source data is collected through the intelligent data acquisition engine. Extract entity elements from the multi-source data and mark each entity element with the timestamp of the corresponding event. The entity elements are associated with the event types in the classification and grading results to generate type association data; Establish a relation topology based on the logical relationships between the entity elements; By integrating the aforementioned types of associated data and relationship topology, an event data association graph containing timestamps is generated.

3. The target event processing method for language model collaboration as described in claim 1, characterized in that, The multi-source data is fused to generate fused data, including: The multi-source data is cleaned to generate cleaned data; The cleaned data is then subjected to format standardization conversion to generate standardized data; Feature vectors are extracted from the standardized data to generate a feature vector set; Based on the feature vector set, correlation mining is performed to generate correlation data; The data relationship data is fused to generate fused data.

4. The target event processing method for language model collaboration as described in claim 1, characterized in that, The fused data is input into a pre-trained language model, and a decision scheme is generated through the pre-trained language model, including: The fused data is subjected to integrity verification to generate verified data; The verified data is mapped to a vector space to generate an input vector. The semantic relationships of the input vector are parsed using a pre-trained language model to generate a semantic relationship graph. The pre-trained language model extracts contextual interaction patterns from the semantic relationship graph to generate contextual features; The pre-trained language model performs cross-domain analogy matching based on the context features to generate an analogy reference set. The policy prototype data is generated by constructing a policy prototype based on the analog reference set using the pre-trained language model. The pre-trained language model is used to optimize the constraints of the policy prototype data to generate a decision scheme.

5. The target event processing method for language model collaboration as described in claim 1, characterized in that, Decision coordination is performed according to the decision scheme, and processing agents are matched and assigned to the decision tasks in the decision scheme, including: Analyze the task elements of the decision-making scheme to generate a decision task set containing all tasks to be assigned; Analyze the execution dependencies of each decision task in the decision task set and generate a task dependency topology graph; Analyze the real-time capability status of currently available processing agents and generate a capability status report that includes load and capability metrics; Based on the task dependency topology graph and the capability status report, dynamic task matching processing is performed to generate a matching result between the decision task and the agent. Based on the matching results between the decision-making task and the agent, optimize the resource allocation strategy and generate the final agent allocation scheme.

6. The target event processing method for language model collaboration as described in claim 1, characterized in that, The processing agent executes a matching decision-making task and generates decision results, including: Based on the agent allocation scheme, task execution instructions for the corresponding decision-making agents are distributed to the corresponding processing agents to generate task instruction sets; The processing agent executes the operation commands in the task instruction set to generate an operation execution record. Monitor the task execution progress and resource status of the processing agent, and generate a status monitoring dataset; Based on the aforementioned status monitoring dataset, abnormal execution behaviors are detected, and an anomaly detection report is generated. By aggregating the operation execution records, status monitoring datasets, and anomaly detection reports, a decision result containing quantitative analysis of the execution effect is generated.

7. The target event processing method for language model collaboration as described in claim 1, characterized in that, After the processing agent executes the matching decision task and generates the decision result, the process further includes: Extract quantitative features of execution effectiveness from the decision results; Based on the quantitative characteristics of the execution effect, analyze the success and failure patterns of decision-making and generate a decision-making pattern analysis report; The decision-making pattern analysis report and the fused data are subjected to spatiotemporal correlation analysis to generate spatiotemporally correlated fused correlation data; Based on the fused and correlated data, language model parameter update data and intelligent agent behavior strategy update data are generated; The inference parameters of the pre-trained language model are updated using the language model parameter update data, and the decision policy library of the processing agent is updated using the processing agent behavior policy update data.

8. A target event processing device for language model collaboration, characterized in that, The target event processing device for language model collaboration includes: The event monitoring module is used to monitor and identify target events and collect multi-source data associated with the target events; The data fusion module is used to fuse the multi-source data to generate fused data; The language model inference module is used to input the fused data into a pre-trained language model and generate a decision scheme through the pre-trained language model. The agent coordination module is used to coordinate decisions according to the decision scheme, and to match and assign processing agents for the decision tasks in the decision scheme. The task execution module is used to execute the matched decision-making task through the processing agent and generate decision results.

9. A computer device, characterized in that, The computer device includes a memory, a processor, and a language model co-operation target event handler stored in the memory and executable on the processor. When the language model co-operation target event handler is executed by the processor, it implements the steps of the language model co-operation target event handling method as described in any one of claims 1-7.

10. A computer-readable storage medium, characterized in that, The storage medium stores a target event handling program for language model collaboration, which, when executed by a processor, implements the steps of the target event handling method for language model collaboration as described in any one of claims 1-7.

Images