A multi-agent based intelligent early warning method, system, device and storage medium

Abstract

This invention proposes a multi-agent-based intelligent early warning method and system. The method includes: acquiring data sources, including IT data and OT data; continuously processing the data sources through a streaming computing framework to generate real-time calculation results, and uniformly storing the real-time calculation results and data sources in a data lake; calling a data governance platform to uniformly manage the data in the data lake, generating a data asset catalog and a field-level lineage graph; using a data monitoring agent to consult the data asset catalog to identify automatically identified key monitoring objects, and combining a prediction model to perform real-time anomaly prediction on the key monitoring objects to obtain monitoring results; and using a task scheduling agent to dynamically optimize task allocation based on the field-level lineage graph, the data asset catalog, and the monitoring results, obtaining and executing the optimized task scheduling strategy, thereby achieving cross-system data connectivity and intelligent collaboration, reducing the cost of manual intervention, and improving the level of intelligence.

Description

Technical Field

[0001] This invention relates to the field of cross-system data processing technology, and in particular to an intelligent early warning method and system based on multiple agents. Background Technology

[0002] In industrial digitalization scenarios, the lack of unified communication protocols and heterogeneous data formats between IT and OT systems leads to difficulties in data integration. Simultaneously, the absence of a comprehensive data governance mechanism results in semantic inconsistencies, low quality, and poor timeliness during data collection, transmission, and application, creating "data breakpoints." This problem hinders cross-system data collaboration and severely impacts the accuracy and timeliness of upper-level applications such as data analysis, anomaly alerts, and task scheduling. Therefore, there is an urgent need for a technical solution that can break down IT / OT barriers, achieve cross-system data semantic alignment, and enable efficient end-to-end collaboration to improve data availability, real-time performance, and business loop capabilities. Summary of the Invention

[0003] This invention provides a multi-agent-based intelligent early warning method and system to address the problems existing in related technologies. The technical solution is as follows:

[0004] In a first aspect, embodiments of the present invention provide a multi-agent-based intelligent early warning method, comprising:

[0005] Acquire data sources, including IT data and OT data; continuously process the data sources through a streaming computing framework to generate real-time computing results, and store the real-time computing results and data sources uniformly in a data lake;

[0006] The data governance platform is invoked to manage the data in the data lake in a unified manner, generating a data asset collection that includes a data asset catalog and a field-level lineage graph.

[0007] The data monitoring agent uses a data asset catalog to identify key monitoring targets that are automatically identified. Then, a predictive model is used to predict anomalies in these key monitoring targets in real time, and the monitoring results are obtained.

[0008] The task scheduling agent dynamically optimizes task allocation based on field-level lineage graphs, data asset catalogs, and monitoring results, and then obtains and executes the optimized task scheduling strategy.

[0009] In one implementation, the continuous processing of IT and OT data using a streaming computing framework to generate real-time computing results includes:

[0010] IT and OT data are sent to a Kafka cluster. The data in the Kafka cluster is partitioned according to the business logic dimension to generate a Kafka topic with multiple partitions. The business logic dimension includes at least one of the following: organization, geographic location, product type, device ID, or timestamp.

[0011] Flink is used to process and analyze data in Kafka topics across different partitions in real time, and output real-time computation results.

[0012] In one implementation, real-time computation results and data sources are uniformly stored in a data lake, including:

[0013] The real-time data stream is filtered to obtain the data that needs to be persistently stored; the real-time data stream includes real-time calculation results and data sources.

[0014] Data that needs to be persistently stored can be directly written to the Delta table through the Delta Lake Sink plugin of the streaming computing framework.

[0015] During the data writing process, partition directories are created based on the specified event time field, and the data to be persistently stored is written to the corresponding time partition path. The columnar storage format Parquet or ORC is used, and the data to be persistently stored is encoded and compressed using the specified compression algorithm before being stored in the data lake.

[0016] In one implementation, invoking a data governance platform to uniformly manage data in a data lake includes:

[0017] Classify and classify the data in the data lake to construct a multi-level data asset catalog;

[0018] Define data standards based on key fields in the multi-level data asset catalog, and unify the data in the data lake into standardized data according to the data standards.

[0019] By combining multi-level data asset catalogs and standardized data, data models and code tables are defined according to business logic;

[0020] By connecting table fields in various data models through data flow logic, a knowledge graph and field-level lineage relationships are formed for customer information, order information, and inventory information;

[0021] Identify key business control nodes based on field-level lineage relationships, establish the association paths between the fields involved in the business control nodes and upstream and downstream data fields, and generate a field-level lineage relationship graph.

[0022] In one implementation, a prediction model is used to perform real-time anomaly prediction on key monitored objects, and the resulting monitoring results include:

[0023] Obtain early warning rules and configure early warning rules for key monitored objects. The configuration of early warning rules includes custom static thresholds and dynamic thresholds. The dynamic thresholds are automatically adjusted based on the distribution characteristics of historical data.

[0024] When the key monitoring object is a single object, a time series prediction model is used to predict anomalies and obtain monitoring results. The time series prediction model includes the ARIMA model or the SARIMA model. The ARIMA model is used to model non-seasonal time series, while the SARIMA model is used to model time series with periodic and seasonal characteristics.

[0025] When there are multiple key monitoring objects in a high-dimensional, multivariable scenario, condition judgments are made based on the data characteristics of the key monitoring objects and business requirements. Based on the condition judgment results, corresponding prediction path branch decisions are determined to predict anomalies and obtain monitoring results. Among them, prediction path branch decisions include nonlinear prediction decisions, linear prediction decisions, supervised learning decisions, or unsupervised learning decisions.

[0026] In one implementation, dynamically optimizing task allocation through a task scheduling agent based on field-level lineage graphs, data asset catalogs, and monitoring results includes:

[0027] Receive task requests from the data monitoring agent, analyze the multidimensional attribute information of the task requests, including time constraints, business priorities, target objects and their associated monitoring results;

[0028] Based on a pre-built field-level lineage graph, the key business control nodes associated with task requests and their upstream and downstream data dependency paths are identified to determine the scope of lineage influence.

[0029] Determine the data asset level of the target object within the data asset system based on the data asset catalog;

[0030] By combining business priority, data asset level, and the scope of influence of lineage, the task priority is dynamically weighted and calculated to generate a comprehensive priority score;

[0031] Task scheduling strategies are dynamically generated based on priority scores, time constraints, and the severity of anomalies in monitoring results.

[0032] In one implementation, the method for executing the optimized task scheduling strategy includes:

[0033] The task type is determined based on the task scheduling strategy, and candidate members are determined based on the task type.

[0034] Obtain the multidimensional role matrix of the candidate members. The role matrix defines the ability dimension and the role ability weight vector for each role.

[0035] Obtain member profile data for candidate members, including professional qualifications, project experience, historical contributions, current workload, and skill mastery.

[0036] The member profile data is mapped to quantitative scores for each ability dimension, forming a member ability vector;

[0037] Calculate the weighted similarity between the member's ability vector and the role's ability weight vector to obtain the ability matching score. Sort the candidate members according to the ability matching score to generate the optimal role allocation scheme.

[0038] Secondly, embodiments of the present invention provide a multi-agent-based intelligent early warning system, comprising:

[0039] The data collection module is used to acquire data sources, including IT data and OT data; it continuously processes the data sources through a streaming computing framework to generate real-time computing results, and stores the real-time computing results and data sources in a unified data lake;

[0040] The data governance module is used to call the data governance platform to manage the data in the data lake in a unified manner and generate a data asset collection that includes a data asset catalog and a field-level lineage graph.

[0041] The data monitoring module is used to identify key monitoring objects automatically by consulting the data asset catalog through the data monitoring agent, and to perform real-time anomaly prediction on the key monitoring objects by combining the prediction model to obtain monitoring results.

[0042] The task scheduling module is used to dynamically optimize task allocation based on field-level lineage graphs, data asset catalogs, and monitoring results through the task scheduling agent, thereby obtaining an optimized task scheduling strategy.

[0043] The task execution module is used to execute the optimized task scheduling strategy.

[0044] Thirdly, embodiments of the present invention provide an electronic device comprising a memory and a processor. The memory and the processor communicate with each other via an internal connection path. The memory stores instructions, and the processor executes the instructions stored in the memory. When the processor executes the instructions stored in the memory, it causes the processor to perform the method described in any of the above embodiments.

[0045] Fourthly, embodiments of the present invention provide a computer-readable storage medium that stores a computer program, wherein when the computer program is run on a computer, the methods in any of the embodiments described above are executed.

[0046] The advantages or beneficial effects of the above technical solutions include at least the following:

[0047] 1. Through intelligent governance, semantic alignment and correlation fusion of accessed IT and OT data are performed to effectively eliminate data ambiguity and conflicts, and improve data consistency and availability;

[0048] 2. By leveraging the collaborative work of the data monitoring agent and the task scheduling agent, a complete automated process from data collection to automatic response is achieved, supporting anomaly prediction based on complex rules and preset models, and significantly improving the accuracy of identifying risks such as hidden equipment faults;

[0049] 3. The task scheduling agent dynamically optimizes resource allocation based on real-time status, avoiding resource overload or idleness, improving execution efficiency by more than 30%, and enhancing the system's adaptability and real-time business response.

[0050] 4. The overall solution enables cross-system data connectivity and intelligent collaboration, reduces the cost of manual intervention, and supports the automation and intelligent upgrading of key businesses such as production and operation and maintenance.

[0051] The above overview is for illustrative purposes only and is not intended to be limiting in any way. In addition to the illustrative aspects, embodiments, and features described above, further aspects, embodiments, and features of the invention will become readily apparent from the accompanying drawings and the following detailed description. Attached Figure Description

[0052] In the accompanying drawings, unless otherwise specified, the same reference numerals throughout the various drawings denote the same or similar parts or elements. These drawings are not necessarily drawn to scale. It should be understood that these drawings depict only some embodiments disclosed in the invention and should not be construed as limiting the scope of the invention.

[0053] Figure 1 This is a flowchart illustrating the intelligent early warning method based on multiple agents of the present invention.

[0054] Figure 2 This is a schematic diagram of the integrated architecture of the present invention;

[0055] Figure 3 This is a schematic diagram of the data governance process of the present invention;

[0056] Figure 4 This is a functional classification diagram of the data monitoring agent of the present invention;

[0057] Figure 5 This is a functional classification diagram of the task scheduling agent of the present invention;

[0058] Figure 6This is a functional classification diagram of the task execution agent of the present invention;

[0059] Figure 7 This is a functional classification diagram of the Task Verification Agent of this invention;

[0060] Figure 8 This is a structural block diagram of an electronic device according to an embodiment of the present invention. Detailed Implementation

[0061] In the following description, only certain exemplary embodiments are briefly described. As those skilled in the art will recognize, the described embodiments can be modified in various ways without departing from the spirit or scope of the invention. Therefore, the drawings and description are considered to be exemplary in nature and not restrictive.

[0062] Example 1

[0063] This embodiment provides a multi-agent-based intelligent early warning method. Through an integrated architecture of "real-time data collaboration + intelligent dynamic decision-making + task self-healing closed loop", it solves the core pain points of traditional solutions such as data fragmentation, delayed response, rigid scheduling, and inefficient collaboration, and realizes intelligent and flexible "collection-monitoring-execution" process.

[0064] Combination Figure 1 , Figure 2 As shown, the intelligent early warning method based on multiple agents in this embodiment specifically includes:

[0065] Step S1: Connect data sources from IT and OT systems, including IT data and OT data; continuously process the data sources through a streaming computing framework to generate real-time computing results, and store the real-time computing results and data sources in a unified data lake.

[0066] It should be noted that IT data includes structured data (databases, APIs) from business systems such as ERP (Enterprise Resource Planning), SRM (Supplier Relationship Management), and PLM (Product Lifecycle Management) within IT systems. OT data, on the other hand, includes real-time data (OPC UA, Modbus, MQTT protocols) from sensors in production equipment, measuring equipment, safety equipment, and environmental protection equipment within OT systems.

[0067] This embodiment uses a streaming computing framework (Apache Flink / Kafka Streams) to process real-time data, combined with batch processing (Spark) to analyze historical data, to build a unified data lake. Specifically:

[0068] Step S11: Send IT data and OT data to the Kafka cluster in a unified manner, and partition the data in the Kafka cluster according to the business logic dimension to generate a Kafka topic with multiple partitions; wherein, the business logic dimension includes at least one of organization, geographical location, product type, device ID or timestamp.

[0069] During data acquisition, partitioning the data according to business logic (such as device ID, production line, order number, etc.) reduces the amount of data in a single partition, avoids data skew issues, and thus improves the parallelism and efficiency of data processing. The partitioned data is then stored in different partitions of a Kafka Topic, allowing consumers to subscribe to specific partitions as needed for efficient data reading and processing.

[0070] Step S12: Use Flink to process and analyze the data in Kafka in real time and output the real-time calculation results.

[0071] In this embodiment, real-time processing refers to cleaning, protocol conversion, timestamp alignment, and dimensional information association (such as associating device IDs with production lines, workshops, product types, etc.) of the raw data streams consumed from each Kafka partition. The purpose is to transform the messy raw data into a structured event stream so that it can be used for subsequent analysis.

[0072] Real-time analysis based on processed data streams, including window aggregation calculations, anomaly detection, trend prediction, and complex event identification, aims to discover anomalies, trends, or business signals from the data and extract valuable information; for example, calculating the average temperature of equipment, detecting abnormal states, and updating order progress.

[0073] Real-time calculation results obtained through real-time processing and analysis include real-time indicators, abnormal alarms, task trigger signals, or prediction information, which are used to drive downstream systems to respond; for example, outputting real-time indicators (such as device ID: D101, current temperature: 78.5℃, status: normal), abnormal alarms (such as alarm type: temperature over-limit, confidence level: 95%, time: 14:30:01), prediction results (device D205, failure probability: 82%, estimated remaining life: 4.2h), etc.

[0074] Enabling a checkpoint mechanism in the streaming computing framework, i.e., enabling Flink's checkpoint mechanism, creates checkpoints periodically. When a Flink job fails, it can recover from the most recent checkpoint, thereby reducing data loss and processing delays. This ensures that the data processing state can be quickly restored in the event of a system failure, guaranteeing production continuity and data accuracy.

[0075] Meanwhile, Flink's backpressure monitoring tool is used to detect pressure points and bottlenecks in the data stream. Based on the monitoring results, the parallelism of the tasks (i.e., the number of tasks running at the same time) is dynamically adjusted, or computing resources (such as CPU, memory, etc.) are increased.

[0076] This embodiment writes the above-mentioned data source and the corresponding real-time calculation results into a data lake for storage. The method for writing the data into the data lake includes:

[0077] Step S13: Filter the real-time data stream to obtain the data that needs to be persistently stored.

[0078] The real-time data stream includes both the source data from the IT / OT system and the real-time computation results processed by Flink, which together constitute a complete data flow chain.

[0079] The real-time data stream is value-classified to identify data that needs to be persistently stored and data that is only used for real-time alerts and does not require long-term analysis.

[0080] Data that is only used for real-time alerts and does not require long-term analysis is marked and intercepted during the stream processing stage, and is not written to the data lake, but only retained as logs for post-event traceability.

[0081] Step S14: The data to be persistently stored is directly written to the Delta table using the Delta Lake Sink plugin of the streaming computing framework. During the data writing process, partition directories are created according to the event time field in hours or days, and the data to be persistently stored is written to the corresponding time partition path, which can speed up subsequent queries. The write operation adopts the columnar storage format Parquet or ORC and enables the Snappy or Zstandard compression algorithm to encode and compress the data before persisting it to the data lake.

[0082] At the same time, historical data is stored in a data lake and divided into hot data and cold data. Hot data is data from the last 7 days and is stored in high-speed SSDs; while cold data is data from more than 30 days and is automatically archived to low-cost storage, such as hard disk drives and tapes.

[0083] Step S2: Call the data governance platform to manage the data in the data lake in a unified manner, and generate a data asset collection that includes a data asset catalog and a field-level lineage graph.

[0084] like Figure 3 As shown, this embodiment uses a data governance platform to classify and grade the data in the data lake, and constructs a multi-level data asset catalog. The multi-level data asset catalog includes L1 business subject domains, L2 business domains, L3 business systems, L4 data tables, and L5 key fields.

[0085] Based on the L5 level key fields in the multi-level data asset catalog, data standards are defined, and data in the data lake is uniformly named, coded, and type-constrained to obtain standardized data.

[0086] By combining multi-level data asset catalogs and standardized data, conceptual, logical, and physical models of the data are defined according to business logic. This process is equivalent to gradually transforming business requirements into a feasible database structure, moving from abstract to concrete. The physical model is implemented based on a relational database or data lake structure, and its fields include code table fields used to identify business status.

[0087] Based on the data flow path and business rule associations between IT and OT systems, graph database technology is used to semantically associate fields in each physical model. Specifically, graph databases such as Neo4j and JanusGraph are used, with "table fields" as nodes and "flow or logical relationships" as edges. All fields and their relationships are organized into a semantic network to construct a knowledge graph between table fields and generate field-level lineage relationships across models and systems.

[0088] In the knowledge graph, a mapping relationship is established between code table fields and corresponding fields in multiple business data tables to identify key business control nodes that affect production scheduling, quality judgment, or equipment control, forming a field-level lineage graph covering the data source to the decision output, so as to support the realization of cross-system business linkage analysis, impact tracing, and automated control objectives.

[0089] Step S3: Based on the data monitoring agent, consult the data asset catalog to determine the key monitoring objects that are automatically identified, and combine the prediction model to perform real-time anomaly prediction on the key monitoring objects to obtain the monitoring results.

[0090] like Figure 4 As shown, in this embodiment, the data monitoring agent performs real-time monitoring and prediction of key monitoring objects according to rules.

[0091] It should be noted that key monitoring objects can be data fields at the L3 level or above in the data asset catalog, and are associated with core business indicators such as production, quality, equipment, safety, or energy consumption.

[0092] Customize early warning rules according to business needs, configure early warning rules for key monitoring objects, and configure early warning rules including custom static thresholds and dynamic thresholds. The dynamic thresholds are automatically adjusted based on the distribution characteristics of historical data.

[0093] When the key monitoring object is a single object, a time-series prediction model is used to predict anomalies and obtain monitoring results. The time-series prediction model includes either an ARIMA model or a SARIMA model. The time-series prediction model primarily feeds and trains the existing model using historical data of the key monitoring object, thereby outputting ideal monitoring results. When the existing model is optimized to an ideal state, it is packaged into an Agent to reduce or even eliminate the AI's hallucination phenomenon. In this embodiment, a crucial periodic change in the time-series prediction function is the season, requiring different models to be trained separately. Details are as follows:

[0094] The ARIMA model is used to model non-seasonal time series. It captures the temporal structure of time series data and makes predictions about current data objects based on historical values ​​and prediction errors on those historical values. ARIMA integrates the autoregressive term (AR) and the moving average term (MA), and can model any non-seasonal time series that exhibits certain patterns.

[0095] If the time series is seasonal, it is modeled using SARIMA (Seasonal ARIMA), which consists of a sequence of time trends and seasonal elements. Predictions are generated by training on historical data and compared with real-time data to identify potential anomalies.

[0096] When there are multiple key monitoring objects in a high-dimensional, multi-variable scenario, high-dimensional real-time data of these objects is acquired. This high-dimensional data includes multi-dimensional variables such as equipment, environment, process, materials, and operations. Based on the data characteristics of the key monitoring objects and business requirements, conditional judgments are made, and the corresponding predictive path branch decision is determined based on the judgment results. Specifically:

[0097] Determine whether high-dimensional data contains labeled results:

[0098] If a label is included, proceed to the supervised learning path;

[0099] If no labels are included, the unsupervised learning path is entered;

[0100] In the supervised learning path, the correlation characteristics between variables are further analyzed. If the relationship is linear, linear discriminant analysis (LDA) or linear regression model is used for prediction. If the relationship is nonlinear, random forest, XGBoost or neural network model is used for prediction.

[0101] In unsupervised learning, if the variables exhibit a linear distribution, Principal Component Analysis (PCA) or Singular Value Decomposition (SVD) is used for dimensionality reduction and anomaly detection. If they exhibit a nonlinear manifold structure, t-SNE, LLE, or CrossCat models are used for clustering and pattern recognition.

[0102] It is important to note that this embodiment prioritizes linear models with low computational overhead for scenarios with high real-time requirements, while nonlinear complex models are selected for scenarios with high accuracy requirements. Joint predictions are performed on key monitored objects based on the selected models, and the output anomaly warnings or trend predictions are used as the monitoring results of the data monitoring agent.

[0103] Step S4: The task scheduling agent dynamically optimizes task allocation based on the field-level lineage graph, data asset catalog, and monitoring results, and obtains and executes the optimized task scheduling strategy.

[0104] like Figure 5 As shown, in this embodiment, the task scheduling agent is the core intelligent component of the task execution layer. It is responsible for dynamically coordinating multi-source task requests (from the data monitoring agent in the early warning and monitoring layer) and achieving efficient and reliable task scheduling through an intelligent decision engine. Its core functions are: analyzing task requirements based on multi-dimensional strategies, completing dynamic task allocation, priority ranking, and execution path planning, and serving as a key hub connecting data processing and task execution.

[0105] This embodiment uses a task scheduling agent to dynamically optimize task allocation based on field-level lineage graphs, data asset catalogs, and monitoring results. The method includes:

[0106] Step S41: Receive a task request from the data monitoring agent. The task request includes task type, target object, execution strategy and priority information.

[0107] Step S42: Perform multi-dimensional strategy analysis on the task request. Based on the results of the multi-dimensional strategy analysis, optimize task scheduling using a dual-dimensional scheduling strategy that combines time and priority dimensions, generating a task scheduling strategy. The multi-dimensional strategy analysis includes:

[0108] Step S421: Obtain multi-dimensional attribute information of the task request. The multi-dimensional attributes include time constraints, business priority, target object, execution resource requirements and related monitoring results.

[0109] Step S422: Based on the pre-built field-level lineage graph, identify the key business control nodes associated with the task request and their upstream and downstream data dependency paths to determine the scope of lineage influence. For example, determine whether it is a prerequisite dependency task or a blocking node in the business process to ensure that the critical path task gets priority to obtain execution resources.

[0110] Step S423: Determine the data asset level of the target object in the data asset system based on the data asset catalog.

[0111] Step S424: Combining the business priority, data asset level, and lineage influence scope, dynamically weight the task priority to generate a comprehensive priority score, assign task urgency weights, and use this to identify high-priority tasks.

[0112] Step S425: Determine whether the task is timed or periodically executed based on the time constraint, and then determine its basic scheduling time window. Based on the comprehensive priority score, the severity and confidence of the abnormality of the monitoring results, dynamically adjust the execution cycle or trigger the emergency queue-jumping mechanism.

[0113] During this process, multiple task requests with the same target object, consistent task type, and common upstream dependency in the lineage graph are identified as tasks that can be merged and batch processing instructions are generated.

[0114] Step S426: The final output includes multi-dimensional analysis results on execution order, resource allocation, merging strategy and scheduling timing, for use in task scheduling decisions.

[0115] During task scheduling, priority can be used as the primary sorting criterion, followed by time, to ensure that high-priority tasks receive execution resources first. When a high-priority task is detected entering the scheduling queue, an emergency queue-jumping mechanism is triggered, pausing the currently executing low-priority task and inserting the high-priority task at the head of the queue for immediate execution. Based on the triggering of high-priority events, the execution cycle of relevant periodic tasks is dynamically adjusted to shorten their scheduling interval and improve response frequency. Through the above collaborative mechanism, joint decision-making based on time and priority dimensions is achieved.

[0116] This embodiment pre-creates different types of task templates using a low-code platform, which are then built into a template library and called upon as needed. Task templates include: 6S scoring, 8D quality reports, balanced scorecards, PFMEA (Process Failure Mode Analysis), and DFMEA (Design Failure Mode Analysis).

[0117] Step S5: Execute the optimized task scheduling strategy through the task execution agent.

[0118] In this embodiment, the task execution agent receives the task scheduling policy, which includes task type, target object, execution time window, priority, and resource requirements. Figure 6 As shown, the task execution agent determines the executors through a task team role and capability matching engine and prepares relevant resources to execute the task, specifically including:

[0119] Based on task type (e.g., R&D, testing, deployment), a multi-dimensional role matrix (technical lead, algorithm engineer, verification specialist, documentation engineer) is obtained, defining the capability dimensions and weight coefficients for each role. Member profile data for each role is acquired, including professional qualifications, project experience, historical contributions, current workload, and skill mastery. This profile data is mapped to quantitative scores for each capability dimension, forming a member capability vector. The weighted similarity between the member capability vector and the role capability weight vector is calculated as a capability matching score, used to evaluate the capability matching degree of candidate members, generating a role-capability mapping graph. Candidate members are ranked according to the matching score to generate the optimal role allocation scheme. The member profile data is updated through historical records and real-time monitoring to ensure the dynamic nature and accuracy of the role-capability mapping.

[0120] At the same time, reinforcement learning models (such as the DQN algorithm) can be used to evaluate task priorities, time windows and member status in real time, dynamically adjust the optimal role allocation scheme, support cross-domain role collaboration (such as technical staff temporarily taking on process coordination roles), and reduce manpower idle rate.

[0121] Furthermore, by analyzing task dependencies through graph neural networks (GNNs), role conflicts (such as resource competition and overlapping responsibilities) are automatically identified, and a secondary negotiation strategy (based on a member willingness model and historical collaboration scores) is triggered. The secondary negotiation strategy considers member willingness and historical collaboration performance to ensure the effectiveness and fairness of conflict resolution.

[0122] After assigning personnel, task instructions and resource configuration information are distributed to the corresponding members, and the task execution status is monitored to form a closed-loop feedback system. The resource configuration information includes a dynamic toolset. The dynamic toolset is generated by integrating RFID IoT technology and a digital twin model. It collects real-time equipment status data (working time, failure rate, calibration cycle) and dynamically generates a minimum necessary toolset based on task process requirements (accuracy level, environmental parameters), avoiding redundant tools. RFID technology is used to track equipment location and status in real time, ensuring timely updates to the toolset. The digital twin model simulates equipment usage in different task scenarios, optimizing tool configuration.

[0123] The aforementioned task execution agent, through intelligent role allocation and resource preparation, achieves efficient team collaboration and resource management, thereby improving the reliability and efficiency of task execution.

[0124] Meanwhile, to better understand the task execution status and visualize the tasks, a real-time digital twin is created for each task, synchronously collecting data on resource consumption, progress deviations, and quality during the execution process, and constructing a closed-loop feedback system of "physical execution - virtual mapping - strategy optimization". A task execution knowledge base is built through indicator knowledge graph technology to provide multi-dimensional decision support such as real-time status monitoring, risk warning (such as critical path delays and resource shortages), and solution simulation (such as "if the person in charge is changed" or "if spare parts are added"), assisting in rapid human intervention in complex scenarios.

[0125] Step S6: The Task Review Agent automatically generates improvement suggestions for completed, evaluated, and summarized tasks.

[0126] like Figure 7 As shown, this embodiment also includes the invocation of a task review agent. The task review agent provides a function based on multi-dimensional intelligent analysis. By constructing a closed-loop task evaluation-summary-optimization mechanism, it achieves quantitative evaluation of task execution effectiveness, experience accumulation, and continuous improvement. The core functions of the task review agent include result evaluation, experience summarization, and improvement suggestion generation. The corresponding modules of each function achieve data flow and collaborative work through an intelligent decision engine.

[0127] The results evaluation module is used to dynamically match and analyze the benefits of completed tasks, including an intelligent evaluation submodule for goal achievement and an efficiency and benefit analysis submodule; as detailed below:

[0128] Step S61: The intelligent evaluation submodule for goal achievement obtains the task description text of the completed goal task, the task template associated with the goal task, and the evaluation criteria.

[0129] The task template refers to a predefined structured framework for tasks, which includes metadata such as task type, objective description, execution process, required roles, and resource types. Evaluation criteria are pre-recorded in a performance evaluation rule base associated with the task type, defining "what constitutes a successful outcome."

[0130] Step S62: The intelligent evaluation submodule for goal achievement uses natural language processing technology to perform semantic analysis on the task description text, extracting key entities, goal intent, and time constraints. Combined with the process structure in the task template and the performance rules in the evaluation criteria, it automatically generates a set of key performance indicators (KPIs). The KPI set includes information such as time completion rate, quality compliance rate, and resource consumption threshold.

[0131] Step S63: The intelligent evaluation submodule for target achievement obtains the actual execution data of the target task, including response time, completion time, result quality, resource consumption, and business impact indicators. The fuzzy comprehensive evaluation method is used to conduct multi-dimensional comprehensive scoring on the actual execution data and the set of key performance indicators, and a deviation analysis report of task execution is generated.

[0132] In this embodiment, the intelligent evaluation submodule for goal achievement is based on the task template used initially and, combined with relevant evaluation criteria (analyzing the task description document using Natural Language Processing (NLP) technology), automatically decomposes it into a set of quantifiable Key Performance Indicators (KPIs) (such as time completion rate, quality compliance rate, resource consumption threshold, etc.). It then employs fuzzy comprehensive evaluation and dynamic threshold calibration technology to dynamically match actual task execution data (such as real-time logs, deliverables, and anomaly records) with preset KPIs, generating a deviation analysis report. Specifically, for unstructured goals (such as creative tasks), a semantic similarity model (such as BERT-BiLSTM) is introduced for qualitative evaluation vectorization.

[0133] The dynamic threshold calibration technology specifically includes:

[0134] Obtain multi-dimensional feature information of the target task, including equipment type, fault code, environmental parameters, task type and resource constraints;

[0135] Based on multidimensional feature information, a feature vector of the target task is constructed, and similar tasks are retrieved from a pre-built historical task library to obtain a group of similar tasks.

[0136] Cluster analysis is performed on similar task groups to determine the complexity level of the target task;

[0137] Based on historical task execution data corresponding to this complexity level, the actual distribution of Key Performance Indicators (KPIs) is statistically analyzed to generate dynamic evaluation thresholds. These dynamic evaluation thresholds serve as the benchmark values ​​for KPI evaluation. The achievement of the target task is then evaluated based on these benchmark values, replacing the preset static thresholds. Simultaneously, the execution results and evaluation data of the current task can be fed back into the historical task database for model iterative optimization.

[0138] The efficiency and effectiveness analysis submodule is used to obtain the full-cycle log data of the target task and build a digital twin model of the target task execution timeline based on the log data. This model records the time consumption of each stage, including requirement confirmation, resource allocation, execution implementation, and result verification.

[0139] Based on the task flow structure, the target task is decomposed into multiple dependent activity nodes. The theoretical shortest time is marked for each activity node. The longest path (i.e., the critical path) is found by traversing all paths using the critical path method. The total time of the critical path is equal to the theoretical shortest duration of the target task.

[0140] The actual total time spent on the target task is determined by using a timeline digital twin model, and the Time Cost Performance Index (TCEI) is quantified by calculating the deviation rate between the theoretical shortest duration and the actual time spent.

[0141] A multi-dimensional resource input model (including labor costs, equipment depreciation, capital consumption, time costs, etc.) is established. Resource consumption logs for each stage are collected through a data platform. Activity-Based Costing (ABC) is used to aggregate the costs in the resource consumption logs, resulting in a refined resource input list. Based on this list, the Data Envelopment Analysis (DEA) algorithm dynamically calculates the Return on Investment (ROI) using resource input as input and task quality and business impact as outputs. Based on the time cost performance index and the ROI, resource optimization suggestions are automatically generated. For example, if the time cost performance index is low (representing wasted time), the optimization suggestion is to compress non-critical path stages and increase parallel processing; if the ROI is low (representing resource redundancy), the optimization suggestion is to replace high-cost resources and reduce unnecessary spare parts carrying.

[0142] The experience summary module in this embodiment is used to summarize the experiences gained from completed and evaluated target tasks, automatically extracting and structuring lessons learned. The core functional units of the experience summary module include an intelligent extraction unit for success factors and an intelligent reflection unit for problems and lessons learned.

[0143] Among them, the intelligent extraction of success factors is mainly used to collect differentiated strategies (such as risk response plans, cross-team collaboration mechanisms, and technological innovations) in historical tasks. After clarifying and standardizing the text data, key information (such as strategy name, implementation steps, and effect evaluation) is extracted. Semantic clustering analysis is performed on the differentiated strategies used in historical tasks to form different strategy clusters. Through association rule mining (Apriori algorithm or FP-Growth algorithm), high-frequency effective strategy combinations are mined in the strategy clusters. The identified effective strategy patterns are organized into standardized templates (such as standardized strategy templates such as "dynamic resource scheduling algorithm" and "agile iterative review mechanism"), and a reusable strategy knowledge base is built for reference and application in subsequent tasks.

[0144] Furthermore, intelligent extraction of success factors also includes collecting multi-source data such as communication records, task allocation logs, and conflict resolution node data generated during collaboration. Based on these data, entities (such as personnel, tasks, and departments) and relationships (such as communication, collaboration, and conflict) are defined. These entities and relationships are then stored and managed using a graph database (such as Neo4j), constructing a knowledge graph of collaboration relationships. Based on this knowledge graph, the contribution of each role (such as information transmission efficiency and decision response speed) is quantified. Efficient collaboration patterns (such as cross-departmental rapid alignment mechanisms and the application effects of real-time visualization collaboration platforms) are analyzed, key features of these patterns are extracted, and best practice guidelines are developed. For example, the "cross-departmental rapid alignment mechanism" is identified as an efficient collaboration pattern, and the application of the "real-time visualization collaboration platform" is found to significantly improve information transparency and decision-making efficiency.

[0145] The Problem and Lessons Learned Intelligent Reflection Unit primarily collects multi-dimensional data related to task deviations (such as delays, rework, and target deviations). It employs the 5 Whys analysis method to delve deeper into the root causes of problems, combining it with Fault Tree Analysis (FTA) technology to construct a fault tree model and clarify the logical relationships between causes at each level. Each failure case is transformed into a feature vector, encompassing multiple dimensions such as technical bottlenecks, process loopholes, and human error. A Siamese neural network is used to compare similar failure patterns, identify common problems, and automatically generate a root cause analysis report based on the analysis results. This report includes categorization tags such as technical bottlenecks, process loopholes, and human error, and provides improvement suggestions to guide subsequent tasks in avoiding similar problems.

[0146] The Problem and Lessons Learned Intelligent Reflection Unit also collects historical successful improvement cases to build a countermeasures library. Each case is analyzed in detail, extracting key elements (such as background, implementation steps, and effect evaluation). Through reinforcement learning (Q-Learning algorithm), it dynamically recommends the optimal combination of improvement strategies based on the characteristics of the current failure scenario (such as resource shortage type and time urgency). For specific failure scenarios, it matches the most suitable solution from the countermeasures library, dynamically adjusts the strategy combination to ensure it adapts to the current environment and conditions, and automatically generates detailed improvement reports, including specific operational steps, expected effects, and potential risks. For example, for the problem of "delay caused by requirement changes," it automatically matches the "requirement change impact assessment model" and the "task priority dynamic adjustment algorithm," providing detailed implementation steps and expected effect evaluations to help the team quickly resolve the problem.

[0147] Through the above steps, the intelligent knowledge extraction system can effectively extract successful experiences and lessons learned from completed and evaluated tasks, forming a structured knowledge base to support the optimization and improvement of subsequent tasks.

[0148] The improvement suggestion generation module is used to automatically generate improvement suggestions for completed, evaluated, and summarized tasks, mainly including two implementation paths:

[0149] Implementation Path 1: The short-term rapid improvement strategy generation unit dynamically adjusts task parameters (such as resource allocation weights and time buffer thresholds) based on real-time deviation data during task execution through reinforcement learning, and generates fine-tuning implementation plans (such as adding a backup resource pool in critical path links and reducing the consumption of resources and time in non-core links).

[0150] Implementation Path Two: The long-term strategic planning simulation unit combines historical data with trend prediction models (LSTM time series forecasting) and task technology stack analysis to formulate a phased technology upgrade plan (such as introducing digital twin simulation technology within 3 years and achieving full-process AI autonomy within 5 years); and based on the team capability matrix and task demand gap analysis, it constructs a talent training intelligent planning model and outputs customized training programs (such as offering specialized courses for "resource scheduling optimization" capabilities and establishing an internal technology sharing mechanism); it extracts cross-task general best practices and constructs a standardized task management process framework (SMPF) through formal methods (Z language) to institutionalize and standardize experience reuse.

[0151] The aforementioned task review agent generates tasks through intelligent task evaluation, experience summarization, and improvement suggestions, enabling quantitative evaluation of task execution effectiveness, experience accumulation, and continuous improvement, thereby enhancing the reliability and efficiency of task execution.

[0152] This embodiment solves the core pain points of traditional solutions, such as data fragmentation, delayed response, rigid scheduling, and inefficient collaboration, through an integrated architecture of "real-time data collaboration + intelligent dynamic decision-making + task self-healing closed loop". It achieves intelligent and flexible "collection-monitoring-execution" process.

[0153] Example 2

[0154] This embodiment provides a multi-agent-based intelligent early warning system, which executes the method described in Embodiment 1 above. The intelligent early warning system includes:

[0155] The data collection module is used to acquire data sources, including IT data and OT data; continuously process the data sources through a streaming computing framework to generate real-time computing results, and store the real-time computing results and the data sources in a unified data lake;

[0156] The data governance module is used to call the data governance platform to manage the data in the data lake in a unified manner and generate a data asset set that includes a data asset catalog and a field-level lineage graph.

[0157] The data monitoring module is used to identify key monitoring objects automatically by consulting the data asset catalog through the data monitoring agent, and to perform real-time anomaly prediction on the key monitoring objects by combining the prediction model to obtain monitoring results.

[0158] The task scheduling module is used to dynamically optimize task allocation based on the field-level lineage graph, the data asset catalog, and the monitoring results through the task scheduling agent, and obtain an optimized task scheduling strategy.

[0159] The task execution module is used to execute the optimized task scheduling strategy through the task execution agent;

[0160] The task review module is used by the task review agent to automatically generate improvement suggestions for completed, evaluated and summarized tasks.

[0161] It should be noted that the functions of each module in this embodiment can be found in the corresponding descriptions in the above methods, and will not be repeated here.

[0162] This embodiment of the system overcomes the technical bottlenecks of traditional IT (Information Technology) and OT (Operational Technology) systems, such as data silos, insufficient real-time collaboration, and rudimentary task scheduling, by constructing an integrated architecture of "data source access - intelligent early warning - task execution." It achieves a full-link intelligent upgrade from data collection to business closed-loop. The following details this from the perspectives of technological innovation, performance improvement, and application value:

[0163] I. Systemic Breakthroughs Brought About by Innovation in Core Technology Architecture

[0164] 1. Leap in heterogeneous data fusion and real-time processing capabilities

[0165] To address the differences in data formats and communication protocols between IT systems (such as ERP and MES) and OT systems (such as PLC and sensors), the data collection module is designed with a multi-source heterogeneous data real-time synchronization access mechanism (data source access layer). Through standardized data interfaces and protocol conversion modules, millisecond-level data interoperability is achieved (traditional solutions have a latency of ≥500ms, while the actual measured latency of this invention is ≤50ms).

[0166] The data governance module cleans, deduplicates, and performs correlation modeling on data through a layered data governance framework, improving data integrity and consistency from 75% in traditional solutions to 99.2%, providing a highly reliable data foundation for upper-level intelligent decision-making.

[0167] 2. Intelligent Agent Clusters Build Dynamic Collaborative Ecosystems

[0168] In this embodiment, the data monitoring module, task scheduling module, task execution module, and task review module form a four-level intelligent agent architecture through data monitoring agent, task scheduling agent, task execution agent, and task review agent, respectively, to achieve a closed loop of "perception-analysis-execution-review". The specific effects are as follows:

[0169] The data monitoring agent, based on a dynamic rule engine (supporting 200+ custom threshold configurations), enables real-time identification of 16 scenarios such as device anomalies and process bottlenecks, with a 40% improvement in early warning accuracy compared to manually set rules.

[0170] The task scheduling agent uses reinforcement learning algorithms to dynamically allocate task priorities. Compared with traditional fixed-rule scheduling, the task conflict rate is reduced by 65% ​​and the resource utilization rate is increased by 32%.

[0171] The task execution agent, combined with the task template library (supporting 20+ configurable templates), enables automated parsing of operation steps and automated preparation of personnel and material lists, thereby accelerating task execution efficiency.

[0172] The task review agent achieves quantitative evaluation of task execution effectiveness, experience accumulation, and continuous improvement by constructing a closed-loop task evaluation-summary-optimization mechanism. The system architecture adopts a layered and modular design, with core functional modules including an outcome evaluation module, an experience summary module, and an improvement suggestion generation module. Each module achieves data flow and collaborative work through an intelligent decision engine.

[0173] 3. End-to-end closed-loop task processing workflow reengineering

[0174] Unlike the traditional fragmented model of "data collection - manual decision-making - offline execution," this embodiment constructs a fully closed-loop process of "real-time monitoring - intelligent decision-making - automatic execution - effect verification." Through a task verification agent, the execution results are verified from multiple dimensions (such as equipment status feedback and process parameter comparison), increasing the task execution success rate from 82% in the traditional approach to 96.5%, and shortening the anomaly handling cycle from an average of 4 hours to 30 minutes.

[0175] The system architecture design in this embodiment adopts a modular plug-in mechanism, which can be seamlessly adapted to various scenarios such as discrete manufacturing, process industries, and smart cities. For example, when deployed in a chemical enterprise, by expanding the OT data access protocol (adding 5 protocols such as Modbus / TCP and OPC UA), the production process can be transformed from "experience-driven" to "data-driven", and the process parameter optimization cycle can be shortened from 72 hours to 4 hours.

[0176] Furthermore, the system in this embodiment can reserve expansion interfaces for edge computing nodes and cloud collaboration channels, supporting future integration with digital twins and industrial internet platforms. Real-world testing data shows that with the proportion of localized processing at the edge increasing to 70%, cloud data interaction volume is reduced by 40%, meeting the industrial internet's security requirement of "data not leaving the factory."

[0177] This embodiment of the technology achieves a triple breakthrough through "data fusion innovation + intelligent agent collaboration + closed-loop process reengineering," solving the core pain points of "data silos, delayed decision-making, and extensive execution" in traditional industrial systems, and realizing a full-chain upgrade from technical architecture to business value. Experimental and application data show that this invention achieves industry-leading levels in key indicators such as data processing efficiency, task execution reliability, and system operation and maintenance costs, demonstrating technical feasibility and commercial value for large-scale promotion.

[0178] Example 3

[0179] This embodiment provides an electronic device. Figure 8 A structural block diagram of an electronic device according to an embodiment of the present invention is shown. Figure 8 As shown, the electronic device includes a memory 100 and a processor 200. The memory 100 stores a computer program that can run on the processor 200. When the processor 200 executes the computer program, it implements the multi-agent-based intelligent early warning method described in the above embodiments. The number of memories 100 and processors 200 can be one or more.

[0180] The electronic device also includes:

[0181] The communication interface 300 is used to communicate with external devices and perform data exchange and transmission.

[0182] If the memory 100, processor 200, and communication interface 300 are implemented independently, they can be interconnected via a bus to communicate with each other. This bus can be an Industry Standard Architecture (ISA) bus, a Peripheral Component Interconnect (PCI) bus, or an Extended Industry Standard Architecture (EISA) bus, etc. This bus can be divided into an address bus, a data bus, a control bus, etc.

[0183] Optionally, in a specific implementation, if the memory 100, processor 200, and communication interface 300 are integrated on a single chip, then the memory 100, processor 200, and communication interface 300 can communicate with each other through an internal interface.

[0184] This invention provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the method provided in this invention.

[0185] This invention also provides a chip, which includes a processor for calling and executing instructions stored in a memory, causing a communication device on which the chip is installed to perform the method provided in this invention.

[0186] This invention also provides a chip, including: an input interface, an output interface, a processor, and a memory. The input interface, output interface, processor, and memory are connected through an internal connection path. The processor is used to execute code in the memory. When the code is executed, the processor is used to execute the method provided in this invention.

[0187] It should be understood that the aforementioned processor can be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. General-purpose processors can be microprocessors or any conventional processor. It is worth noting that the processor can be a processor supporting the Advanced Reduced Instruction Set Computing (RISC) machine (ARM) architecture.

[0188] Further, optionally, the aforementioned memory may include read-only memory and random access memory, and may also include non-volatile random access memory. The memory may be volatile or non-volatile, or may include both. Non-volatile memory may include read-only memory (ROM), programmable read-only memory (PROM), erasable programmable read-only memory (EPROM), electrically erasable programmable read-only memory (EEPROM), or flash memory. Volatile memory may include random access memory (RAM), which serves as an external cache. Many forms of RAM are available by way of example, but not limitation. Examples include static random access memory (SRAM), dynamic random access memory (DRAM), synchronous dynamic random access memory (SDRAM), double data rate synchronous dynamic random access memory (DDR SDRAM), enhanced synchronous dynamic random access memory (ESDRAM), synchronous linked dynamic random access memory (SLDRAM), and direct rambus RAM (DR RAM).

[0189] In the above embodiments, implementation can be achieved, in whole or in part, by software, hardware, firmware, or any combination thereof. When implemented in software, it can be implemented, in whole or in part, as a computer program product. A computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, all or part of the flow or function according to the present invention is generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transferred from one computer-readable storage medium to another.

[0190] In the description of this specification, references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the invention. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Moreover, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of those different embodiments or examples.

[0191] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include at least one of that feature. In the description of this invention, "a plurality of" means two or more, unless otherwise explicitly specified.

[0192] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any person skilled in the art can easily conceive of various variations or substitutions within the technical scope disclosed in the present invention, and these should all be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A multi-agent based intelligent early warning method, characterized in that, include: Obtain data sources, including IT data and OT data; The data source is continuously processed through a streaming computing framework to generate real-time computing results, and the real-time computing results and the data source are stored in a unified data lake. The data governance platform is invoked to manage the data in the data lake in a unified manner, generating a data asset set that includes a data asset catalog and a field-level lineage graph. Based on the data monitoring agent, the key monitoring objects are identified by consulting the data asset catalog. The key monitoring objects are then used to make real-time anomaly predictions in conjunction with the prediction model to obtain the monitoring results. The task scheduling agent dynamically optimizes task allocation based on the field-level lineage graph, the data asset catalog, and the monitoring results, thereby obtaining and executing the optimized task scheduling strategy. This dynamic optimization of task allocation based on the field-level lineage graph, the data asset catalog, and the monitoring results includes: Receive a task request from the data monitoring agent, and analyze the multi-dimensional attribute information of the task request, including time constraints, business priority, target object and its associated monitoring results; Based on a pre-built field-level lineage graph, the key business control nodes associated with the task request and their upstream and downstream data dependency paths are identified to determine the scope of lineage influence. The data asset level of the target object in the data asset system is determined according to the data asset catalog; By combining the aforementioned business priorities, data asset levels, and the scope of influence of lineage, the task priorities are dynamically weighted and calculated to generate a comprehensive priority score. A task scheduling strategy is dynamically generated based on the priority score, the time constraint, and the severity of the anomalies in the monitoring results.

2. The multi-agent based intelligent early warning method according to claim 1, characterized in that, The continuous processing of the IT data and OT data using a streaming computing framework to generate real-time computing results includes: The IT data and OT data are uniformly sent to the Kafka cluster. The data in the Kafka cluster is partitioned according to the business logic dimension to generate a Kafka topic with multiple partitions. The business logic dimension includes at least one of organization, geographical location, product type, device ID or timestamp. Flink is used to process and analyze the data in each partitioned Kafka topic in real time and output real-time calculation results.

3. The intelligent early warning method based on multiple agents according to claim 1, characterized in that, The real-time calculation results and the data source are uniformly stored in the data lake, including: The real-time data stream is filtered to obtain the data that needs to be persistently stored; wherein the real-time data stream includes the real-time calculation results and the data source; Data that needs to be persistently stored can be directly written to the Delta table through the Delta Lake Sink plugin of the streaming computing framework. During the data writing process, partition directories are created based on the specified event time field, and the data to be persistently stored is written to the corresponding time partition path. The columnar storage format Parquet or ORC is used, and the data to be persistently stored is encoded and compressed using the specified compression algorithm before being stored in the data lake.

4. The intelligent early warning method based on multiple agents according to claim 1, characterized in that, The process of calling the data governance platform to uniformly manage the data in the data lake and generating a data asset set including a data asset catalog and a field-level lineage graph includes: The data in the data lake is classified and graded to construct a multi-level data asset catalog; Data standards are defined according to the key fields in the multi-level data asset catalog, and the data in the data lake is unified into standardized data according to the data standards. Based on the multi-level data asset catalog and the standardized data, define the data model and code table according to business logic; By connecting table fields in various data models through data flow logic, a knowledge graph and field-level lineage relationships are formed for customer information, order information, and inventory information; Identify key business control nodes based on field-level lineage relationships, establish the association paths between the fields involved in the business control nodes and upstream and downstream data fields, and generate a field-level lineage relationship graph.

5. The intelligent early warning method based on multiple agents according to claim 1, characterized in that, The real-time anomaly prediction of the key monitored objects using the combined prediction model yields the following monitoring results: When the key monitoring object is a single object, an anomaly prediction model is used to obtain the monitoring results; the time series prediction model includes an ARIMA model or a SARIMA model, wherein the ARIMA model is used to model non-seasonal time series, and the SARIMA model is used to model time series with periodic and seasonal characteristics. When there are multiple key monitoring objects and they are in a high-dimensional, multivariate scenario, condition judgments are made based on the data characteristics of the key monitoring objects and business requirements. Based on the condition judgment results, corresponding prediction path branch decisions are determined to perform anomaly prediction and obtain monitoring results. Among them, prediction path branch decisions include nonlinear prediction decisions, linear prediction decisions, supervised learning decisions, or unsupervised learning decisions.

6. The intelligent early warning method based on multiple agents according to claim 1, characterized in that, Methods for implementing optimized task scheduling strategies include: The task type is determined based on the task scheduling strategy, and candidate members are determined based on the task type. Obtain the multidimensional role matrix of the candidate members, wherein the role matrix defines the ability dimension and the role ability weight vector corresponding to each role; Obtain the member profile data of the candidate members, which includes professional qualifications, project experience, historical contributions, current workload, and skill mastery. The member profile data is mapped to quantitative scores for each ability dimension to form a member ability vector; Calculate the weighted similarity between the member's ability vector and the role's ability weight vector to obtain the ability matching score. Sort the candidate members according to the ability matching score to generate the optimal role allocation scheme.

7. A multi-agent-based intelligent early warning system, characterized in that, include: The data collection module is used to acquire data sources, including IT data and OT data; The data source is continuously processed through a streaming computing framework to generate real-time computing results, and the real-time computing results and the data source are stored in a unified data lake. The data governance module is used to call the data governance platform to manage the data in the data lake in a unified manner and generate a data asset set that includes a data asset catalog and a field-level lineage graph. The data monitoring module is used to identify key monitoring objects automatically by consulting the data asset catalog through the data monitoring agent, and to perform real-time anomaly prediction on the key monitoring objects by combining the prediction model to obtain monitoring results. The task scheduling module is used to dynamically optimize task allocation based on the field-level lineage graph, the data asset catalog, and the monitoring results using a task scheduling agent, thereby obtaining an optimized task scheduling strategy. The dynamic optimization of task allocation based on the field-level lineage graph, the data asset catalog, and the monitoring results using the task scheduling agent includes: Receive a task request from the data monitoring agent, and analyze the multi-dimensional attribute information of the task request, including time constraints, business priority, target object and its associated monitoring results; Based on a pre-built field-level lineage graph, the key business control nodes associated with the task request and their upstream and downstream data dependency paths are identified to determine the scope of lineage influence. The data asset level of the target object in the data asset system is determined according to the data asset catalog; By combining the aforementioned business priorities, data asset levels, and the scope of influence of lineage, the task priorities are dynamically weighted and calculated to generate a comprehensive priority score. A task scheduling strategy is dynamically generated based on the priority score, the time constraint, and the severity of the anomalies in the monitoring results. The task execution module is used to execute optimized task scheduling strategies through the task execution agent.

8. An electronic device, characterized in that, include: A processor and a memory, wherein the memory stores instructions that are loaded and executed by the processor to implement the multi-agent-based intelligent early warning method as described in any one of claims 1 to 6.

9. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program, which, when executed by a processor, implements the multi-agent-based intelligent early warning method as described in any one of claims 1 to 6.

Images