An agent scheduling method and device

Abstract

Embodiments of the present application provide an agent scheduling method and device, which belong to the technical field of computers. The method comprises: obtaining an input of a user, the type of the input including text, voice and / or image; pre-processing the input to obtain model input information; based on a large language model, analyzing the model input information to obtain a task instruction, the task instruction including a task type, parameter information and a dependency relationship of at least one subtask; determining an execution order of the subtasks based on the dependency relationship; assigning an agent to each subtask based on the task type of each subtask; and scheduling the assigned agents based on the execution order and the parameter information. The method is used to solve at least one problem existing in the prior art.

Description

Technical Field

[0001] This invention relates to the field of computer technology, and more specifically to an intelligent agent scheduling method and device. Background Technology

[0002] In the digital transformation of the petroleum industry, various intelligent agents, such as data analysis agents, report generation agents, and visualization agents, are widely used in various aspects of exploration and development, production management, and decision support. However, existing methods for scheduling and orchestrating tasks for intelligent agents still have the following limitations: (1) Manual triggering is the main method: Most intelligent agents need to be manually triggered to execute and cannot be automatically started according to actual needs. In the petroleum industry, many routine tasks such as daily report generation and production data statistics still require manual operation, which is inefficient and prone to errors.

[0003] (2) Long waiting time: When performing time-consuming tasks, such as generating complex reports, large PPTs, or 3D geological models, users need to wait for a long time. For example, generating a complete oil and gas field development plan may take several hours or even days, during which time users cannot obtain any progress information, affecting work efficiency.

[0004] (3) Low resource utilization: The lack of collaborative working mechanisms among agents leads to a waste of computing resources. In the petroleum industry, different departments often deploy and use agents independently, resulting in redundant resource allocation and idle resources.

[0005] (4) Complex task orchestration: When multiple agents work together to complete complex tasks, the orchestration process is cumbersome and prone to errors. For example, numerical simulation of oil and gas reservoirs requires the collaborative work of multiple steps and multiple agents. Manual orchestration is not only complex but also difficult to optimize.

[0006] (5) Difficulty in previewing results: The task execution process is not transparent, and users cannot understand the task progress and intermediate results in a timely manner. When performing large-scale data analysis tasks, users often cannot know whether the task is progressing normally or whether parameters need to be adjusted.

[0007] Based on the aforementioned problems, existing technological solutions are insufficient to meet the petroleum industry's demands for efficient, flexible, and intelligent agent scheduling and task orchestration. The limitations of existing methods become even more apparent when facing complex tasks and dynamically changing business requirements. Summary of the Invention

[0008] The purpose of this invention is to provide an agent scheduling method and device to solve at least one problem existing in the prior art.

[0009] To achieve the above objectives, in a first aspect, embodiments of the present invention provide an agent scheduling method, the method comprising: acquiring user input, the type of which includes text, speech, and / or images; preprocessing the input to obtain model input information; analyzing the model input information based on a large language model to obtain task instructions, the task instructions including task type, parameter information, and dependencies of at least one subtask; determining the execution order of the subtasks based on the dependencies; assigning an agent to each subtask based on the task type of each subtask; and scheduling the assigned agents based on the execution order and parameter information.

[0010] Optionally, preprocessing the input to obtain model input information includes: converting each type of input into a feature vector based on an encoder, wherein the encoder includes a text encoder, a speech encoder, and / or an image encoder; if the input type is a single type, then the converted feature vector is used as the model input information; and if the input type is greater than or equal to two, then the converted multiple feature vectors are fused across modally, and the fused multimodal feature vector is used as the model input information.

[0011] Optionally, based on a large language model, analyzing the model input information to obtain task instructions includes: performing intent recognition on the model input information to obtain task intent; decomposing the task intent to obtain at least one subtask and its task type, wherein the task type includes data statistics, data fusion, report generation, data analysis, 3D modeling, numerical simulation, and / or visualization; extracting parameter information for each subtask from the model input information, wherein the parameter information includes oil and gas reservoir name, block range, time period, data type, output format, and / or industry standard requirements; completing the parameter information for each subtask based on a pre-constructed knowledge graph; analyzing each subtask after completing the parameter information to determine the input and output data of each subtask; and analyzing the subtasks based on the input and output data to obtain the dependencies between the subtasks.

[0012] Optionally, the intent recognition of the model input information further includes: detecting the task intent, determining whether there are any abnormal issues in the task intent, the abnormal issues including ambiguity, conflict, or error; and if so, performing exception handling until there are no abnormal issues in the task intent, the exception handling including: prompting error information to the user; receiving new input from the user; and performing intent recognition based on context awareness and the new input from the user to obtain a new task intent.

[0013] Optionally, the task instruction further includes the priority of the subtasks, and the method further includes: after determining the execution order of the subtasks, determining whether there are subtasks that are executed in parallel; and if so, sorting the subtasks that are executed in parallel based on the priority of each subtask to obtain a new execution order of the subtasks, wherein the priority includes low priority, medium priority and high priority.

[0014] Optionally, assigning an agent to each subtask based on its task type includes: performing the following processing on each subtask: matching at least one candidate agent based on the task type of the subtask; obtaining state information for each candidate agent, including current load, historical execution data, resource usage, health status, and / or functional adaptability; calculating a multi-dimensional score for each candidate agent based on the state information; and selecting the candidate agent with the highest multi-dimensional score as the optimal agent and assigning the optimal agent to the subtask.

[0015] Optionally, scheduling the allocated agents based on the execution order and parameter information includes: starting the agent corresponding to each subtask to execute the subtask based on the execution order and parameter information, and monitoring the resource utilization and faults of each agent in real time during execution. The resource utilization includes the agent's CPU utilization, memory usage, and / or storage usage. Based on the resource utilization, determining whether the agent is overloaded. If any agent is overloaded, checking if there are any subtasks to be executed for that agent. If so, transferring the subtasks to be executed to a backup agent, which includes the agents in the candidate agents. If any agent fails, transferring all of its subtasks to the backup agent.

[0016] Optionally, the method further includes: collecting resource utilization and fault issues of each agent during execution; and updating the historical execution data based on the collected resource utilization and fault issues.

[0017] Optionally, when the intelligent agent is a data acquisition intelligent agent, the execution of the sub-task includes: connecting to a data source and acquiring the required data based on the parameter information, the data source including an oilfield SCADA system or a database; preprocessing the acquired data, the preprocessing including data cleaning and data verification; submitting the preprocessed data to a preset path for storage; detecting the integrity of the preprocessed data, and if missing data is found, adding a missing data completion sub-task, the missing data completion sub-task containing detailed information about the missing data, the detailed information including well number, time range and / or missing type; setting the missing data completion sub-task as a high priority and assigning it to the data repair intelligent agent for execution; and finding sub-tasks that depend on the data acquisition intelligent agent, pausing their execution until the data repair intelligent agent completes its execution.

[0018] Optionally, the data repair agent performing the missing data completion subtask includes: determining the missing time of the missing data based on the detailed information of the missing data; completing the missing data based on the missing time to obtain repaired data; verifying the rationality of the repaired data based on historical data and / or adjacent well data; and if the rationality verification passes, generating a repair report and submitting the repaired data and the repair report to a preset path for storage, wherein the repair report includes a summary of the repaired data and the rationality verification result.

[0019] Optionally, the missing data completion based on the missing time includes: if the missing time is less than a first time threshold, then using linear interpolation to complete the missing data; if the missing time is greater than or equal to the first time threshold and less than a second time threshold, then combining historical data from the same period and adjacent well data to predictively complete the missing data; and if the missing time is greater than or equal to the second time threshold, then using a pre-trained prediction model to predictively complete the missing data.

[0020] Optionally, the method further includes: when scheduling the assigned intelligent agents, tracking the execution progress and intermediate results of each subtask in real time, and transmitting the intermediate results and execution progress to the user terminal for real-time preview; and after scheduling is completed, integrating and outputting the execution results of all subtasks.

[0021] On the other hand, embodiments of the present invention provide a device for intelligent agent scheduling, the device including a memory and a processor, the processor being used to run a program, wherein the program, when run, is used to execute any of the methods described above.

[0022] The method proposed in this invention leverages the semantic understanding capabilities of large language models to achieve deep understanding and parsing of user natural language input, accurately grasping user needs and adapting to complex and diverse business scenarios in the petroleum industry, such as exploration and development data analysis, production report generation, and reservoir simulation parameter configuration. After analyzing and obtaining task instructions, the method proposed in this invention can automatically complete task scheduling, orchestration, and execution, reducing user intervention, significantly lowering user operational complexity, and improving processing efficiency.

[0023] Other features and advantages of the embodiments of the present invention will be described in detail in the following detailed description section. Attached Figure Description

[0024] The accompanying drawings are provided to further illustrate embodiments of the present invention and form part of the specification. They are used together with the following detailed description to explain the embodiments of the present invention, but do not constitute a limitation thereof. In the drawings: Figure 1 This is a flowchart illustrating the agent scheduling method provided in an embodiment of the present invention; Figure 2 This is a schematic diagram of the device provided in an embodiment of the present invention.

[0025] Explanation of reference numerals in the attached figures 101 Processor 102 Memory 103 bus 10 devices Detailed Implementation The specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. It should be understood that the specific embodiments described herein are for illustration and explanation only and are not intended to limit the scope of the present invention.

[0026] It should be noted that the acquisition, transmission, storage, use, and processing of data in the technical solution of this application all comply with relevant laws and regulations. In the embodiments of this application, certain existing industry solutions such as software, components, and models may be mentioned. These should be considered exemplary, intended only to illustrate the feasibility of implementing the technical solution of this application, and do not imply that the applicant has already used or necessarily used such solutions.

[0027] With the rapid development of artificial intelligence technology, especially the breakthrough in Large Language Model (LLM) technology, new technical approaches have been provided for intelligent scheduling and task orchestration. Large models offer the following advantages: (1) Natural language understanding ability: It can understand the complex needs expressed by users in natural language and reduce the complexity of user operations.

[0028] (2) Knowledge reasoning ability: Possesses strong knowledge reserves and reasoning ability, and is able to handle complex business logic.

[0029] (3) Adaptive learning ability: It can learn from historical data and continuously optimize the scheduling strategy.

[0030] (4) Multimodal processing capability: It can process various types of data such as text, images, and tables.

[0031] (5) Dialogue interaction capability: Supports task interaction and adjustment through dialogue.

[0032] These advantages make large-scale models an ideal technological foundation for building intelligent scheduling and task orchestration systems. Combining knowledge graphs, machine learning, and other technologies, more intelligent, efficient, and flexible agent scheduling and task orchestration methods can be achieved. Based on large-scale language models, this invention proposes an agent scheduling method. Figure 1 This is a flowchart illustrating the agent scheduling method provided in an embodiment of the present invention, as shown below. Figure 1 As shown, the method includes the following steps S1 to S6.

[0033] Step S1: Obtain user input, the type of which includes text, voice and / or images.

[0034] Step S2: Preprocess the input to obtain model input information.

[0035] Step S3: Based on the large language model, analyze the input information of the model to obtain task instructions. The task instructions include the task type, parameter information and dependencies of at least one subtask.

[0036] Step S4: Determine the execution order of the subtasks based on the dependencies.

[0037] Step S5: Assign an agent to each subtask based on its task type.

[0038] Step S6: Based on the execution order and parameter information, schedule the assigned agents.

[0039] The method proposed in this invention analyzes the user's natural language expression based on a large language model, determines the user's task requirements, and transforms the natural language description into machine-executable task instructions, eliminating the need for the user to learn complex commands or interfaces. Upon receiving the task instructions, the method can automatically complete the scheduling, orchestration, and execution of tasks, reducing user intervention, significantly lowering user operational complexity, and improving processing efficiency.

[0040] Furthermore, the input is preprocessed to obtain model input information, including: converting each type of input into a feature vector based on an encoder, wherein the encoder includes a text encoder, a speech encoder, and / or an image encoder; if the input type is a single type, the converted feature vector is used as the model input information; and if the input type is greater than or equal to two, the converted multiple feature vectors are fused across modally, and the fused multimodal feature vector is used as the model input information.

[0041] The method proposed in this invention supports not only text input but also the understanding of multimodal inputs such as voice and images. For example, for a user's voice command "Generate an analysis report on crude oil production in Block B for the past three months," it can first be converted into a voice feature vector, and then this voice feature vector can be used as the model input information for analysis by a large language model. For inputs with two or more types, such as a user uploading a geological image containing well location distribution and the accompanying text "Analyze the correlation of production data at well locations in the image," then text feature vectors and image feature vectors can be obtained based on a text encoder and an image encoder, respectively, and then the two can be fused across modally. The fused multimodal feature vector can then be used as the model input information. After receiving the model input information, the large language model can clearly understand that the task is a correlation analysis of production data based on specific well locations. Existing fusion techniques can be used for the cross-modal fusion of multiple feature vectors, which will not be elaborated upon in this invention.

[0042] Furthermore, based on a large language model, the analysis of the model input information to obtain task instructions includes: performing intent recognition on the model input information to obtain task intent; decomposing the task intent to obtain at least one subtask and its task type, wherein the task type includes data statistics, data fusion, report generation, data analysis, 3D modeling, numerical simulation, and / or visualization; extracting parameter information for each subtask from the model input information, wherein the parameter information includes oil and gas reservoir name, block range, time period, data type, output format, and / or industry standard requirements; completing the parameter information for each subtask based on a pre-constructed knowledge graph; analyzing each subtask after completing the parameter information to determine the input and output data of each subtask; and analyzing the subtasks based on the input and output data to obtain the dependencies between the subtasks.

[0043] By inputting information into the model, the task type can be automatically identified and accurately matched to specific scenarios within the petroleum industry. For example, if a user inputs "statistically calculate the daily production of each oil well in Oilfield E and generate a comparison table," the analysis can identify it as a composite task type of "data statistics + table visualization." If a user inputs "build a 3D geological model based on seismic and well logging data from Block F," the analysis can identify it as a task type of "multi-source data fusion + 3D modeling."

[0044] Pre-built knowledge graphs can store and manage domain knowledge, providing knowledge support for task understanding and scheduling. Using domain knowledge graphs, parameter information required for a task can be extracted and completed from user input. For example, if a user inputs "generate a numerical simulation report for reservoir C," the system automatically completes the default parameters required for the numerical simulation, such as the simulation time step, initial reservoir pressure, and permeability range, using the petroleum industry knowledge graph, while extracting the core parameter "reservoir C." If a user inputs "analyze the reservoir performance of block D," the system can supplement the reservoir performance analysis with key parameter dimensions such as porosity, oil saturation, and effective thickness, based on the knowledge graph.

[0045] Furthermore, the intent recognition of the model input information also includes: detecting the task intent, determining whether the task intent has any abnormal issues, including ambiguity, conflict, or error; and if so, performing exception handling until the task intent has no abnormal issues, the exception handling including: The system displays error messages to the user; receives new input from the user; and performs intent recognition based on context awareness and the new input to obtain a new task intent.

[0046] The method proposed in this invention enables multi-turn dialogue interaction. If an anomaly is detected in the task intent, multiple rounds of dialogue are initiated until a new task intent without anomalies is obtained. During the multi-turn dialogue interaction, the large language model can automatically understand the user's historical preferences and domain knowledge, automatically complete the parameter information required for the task, and have it confirmed by the user, thereby improving the speed of task comprehension.

[0047] In some embodiments of the present invention, when performing intent recognition on the model input information, a judgment is also made as to whether execution should begin. For example, a user first asks for "the annual production data of oil field A," then says "to create a visualization based on this data," and finally says "to begin execution." After receiving the final input, the task intent is then decomposed. For the above input, by associating it with the preceding context, it is determined that the user's intent is to perform visualization analysis of the annual production data of oil field A, rather than two independent tasks, thus ultimately resulting in a sub-task.

[0048] Furthermore, the task instruction also includes the priority of the subtasks, and the method further includes: after determining the execution order of the subtasks, determining whether there are subtasks that are executed in parallel; and if so, sorting the subtasks that are executed in parallel based on the priority of each subtask to obtain a new execution order of the subtasks, wherein the priority includes low priority, medium priority and high priority.

[0049] In practice, the priority of subtasks can be obtained by analyzing the model input information through a large model. For example, the large model performs keyword retrieval on the model input information. When a clear keyword indicating "low priority," "medium priority," or "high priority" is found, the priority is directly set according to the corresponding keyword. When keywords such as "urgent," "immediately," or "right away" are found, the corresponding subtask is marked as high priority. For priorities that cannot be determined by keywords, the priority of subtasks can be directly marked based on a preset task type-priority mapping table.

[0050] The method proposed in this invention enables priority management, supports priority-based task scheduling, and ensures that important tasks are executed first. For tasks that can be executed in parallel, parallel execution is rationally arranged to shorten the overall execution time. For example, when generating production reports for multiple regions, data from different regions can be processed simultaneously, significantly reducing report generation time. In the petroleum industry, priority management ensures that urgent tasks such as well kick handling and analysis are given priority.

[0051] Furthermore, based on the task type of each subtask, assigning an agent to each subtask includes performing the following processing on each subtask: Based on the task type of the subtask, at least one candidate agent is matched; the status information of each candidate agent is obtained, including current load, historical execution data, resource usage, health status and / or functional adaptability; based on the status information, a multi-dimensional score is calculated for each candidate agent; and the candidate agent with the highest multi-dimensional score is selected as the optimal agent, and the optimal agent is assigned to the subtask.

[0052] When assigning agents, candidate agents are first matched based on the task type to improve the success rate of task execution. In some embodiments of the present invention, the selection of candidate agents is also combined with domain knowledge stored in the knowledge graph. For example, when the task is "numerical simulation of high-sulfur gas reservoirs", the system learns from domain knowledge that this type of simulation requires special phase equations and corrosion factor calculations. In this case, a dedicated agent with high-sulfur gas reservoir simulation capabilities is automatically selected, rather than a general oil reservoir simulation agent. When the task is "generating a reserve assessment report that conforms to industry standard SY / T 5329-2019", a report generation agent familiar with the standard's format and parameter requirements is automatically selected.

[0053] After matching candidate agents, a multi-dimensional score is calculated for each agent. For example, based on preset rules, each state information is assigned a value, and then weights are allocated to multiple state information and their sums are calculated to obtain a multi-dimensional score. Finally, the candidate agents are ranked based on the multi-dimensional scores to find the optimal agent. By incorporating current load and resource usage into the scoring criteria, resource utilization can be effectively improved.

[0054] In some embodiments of the present invention, different task types are preset with different combinations of weight values. For example, when performing a large reservoir simulation task, sufficient computing resources need to be reserved in advance to ensure that the task can be completed smoothly. Therefore, for three-dimensional modeling tasks, higher weight values ​​are set for "current load" and "resource usage" to ensure the resources required for task execution and avoid task failure due to insufficient resources.

[0055] Furthermore, scheduling the allocated agents based on the execution order and parameter information includes: starting the agent corresponding to each subtask to execute the subtask based on the execution order and parameter information, and monitoring the resource utilization and faults of each agent in real time during execution. The resource utilization includes the agent's CPU utilization, memory usage, and / or storage usage. Based on the resource utilization, determining whether the agent is overloaded. If any agent is overloaded, checking if there are any subtasks to be executed for that agent. If so, transferring the subtasks to be executed to a backup agent, which includes the agents in the candidate agents. And if any agent fails, transferring all of its subtasks to the backup agent.

[0056] This invention enables real-time monitoring during the execution of subtasks. By determining in real time whether an agent is malfunctioning or overloaded, the subtask can be automatically switched to a backup agent to continue execution. This balances the task load among multiple agents, avoids single-point overload, and ensures the smooth execution of subtasks. This adjustment method ensures the stability of the overall system performance and prevents the overload of a single agent from affecting the operation of the entire system.

[0057] For example, if the system monitors in real time that "Data Analysis Agent A" currently has a CPU utilization of 85% and a memory usage of 70%, then it can be determined that "Data Analysis Agent A" is overloaded. Simultaneously, if "Data Analysis Agent B" is monitored to have a CPU utilization of 30% and a memory usage of 40%, then the subtask "G Block Production Data Trend Analysis" to be executed by "Data Analysis Agent A" will be automatically assigned to "Data Analysis Agent B". If all data analysis agents have high loads (CPU utilization ≥ 75%), then the priority of the subtasks will be used to determine whether the subtasks are urgent, and some non-urgent tasks (such as data backup statistics) will be temporarily stored in the task queue and executed after resources are released.

[0058] In some embodiments of the present invention, if the agent is overloaded, it can be determined whether the subtask to be executed can be split. If it can be split, the multiple subtasks obtained from the split can be transferred to multiple backup agents respectively.

[0059] For example, during the execution of the "H oil and gas reservoir 3D modeling" task, if the agent responsible for seismic data processing is found to have malfunctioned, a backup agent is immediately activated to take over its corresponding subtask, ensuring that the modeling task is not affected.

[0060] Furthermore, the method also includes: collecting resource utilization and fault issues of each agent during execution; and updating the historical execution data based on the collected resource utilization and fault issues.

[0061] By collecting various data and updating historical execution data, scheduling strategies and task orchestration methods can be continuously optimized, thereby continuously improving the accuracy of agent matching.

[0062] Furthermore, when the intelligent agent is a data acquisition intelligent agent, the execution of the sub-task includes: connecting to a data source and acquiring the required data based on the parameter information, the data source including an oilfield SCADA system or a database; preprocessing the acquired data, the preprocessing including data cleaning and data verification; submitting the preprocessed data to a preset path for storage; detecting the integrity of the preprocessed data, and if missing data is found, adding a missing data completion sub-task, the missing data completion sub-task containing detailed information about the missing data, the detailed information including well number, time range and / or missing type; setting the missing data completion sub-task as a high priority and assigning it to the data repair intelligent agent for execution; and finding sub-tasks that depend on the data acquisition intelligent agent, pausing their execution until the data repair intelligent agent completes its execution.

[0063] In some embodiments of the present invention, when missing data is detected, the system first automatically attempts to connect to other data sources to obtain the required data. If the system still cannot obtain complete data, a new subtask for completing the missing data is added.

[0064] Furthermore, the data repair agent performing the missing data completion subtask includes: determining the missing time of the missing data based on the detailed information of the missing data; completing the missing data based on the missing time to obtain repaired data; verifying the rationality of the repaired data based on historical data and / or adjacent well data; and if the rationality verification passes, generating a repair report and submitting the repaired data and the repair report to a preset path for storage, wherein the repair report includes a summary of the repaired data and the rationality verification result.

[0065] Specifically, during the initialization phase, the data repair agent receives task instructions, parses parameters such as the target oil field, data type (e.g., daily production), and time range, then connects to the data source and formulates an initial data acquisition plan, including acquisition frequency and data range. After initialization, it begins data acquisition. When saving data, the preset path can be a preset receiving path in the intelligent scheduling engine, which is responsible for scheduling all agents.

[0066] Furthermore, based on the missing time, the missing data completion includes: if the missing time is less than a first time threshold, then using linear interpolation to complete the missing data; if the missing time is greater than or equal to the first time threshold and less than a second time threshold, then combining historical data from the same period and adjacent well data to predictively complete the missing data; and if the missing time is greater than or equal to the second time threshold, then using a pre-trained prediction model to predictively complete the missing data.

[0067] The pre-trained prediction model can be the Prophet model, or other prediction models can be selected according to the actual situation. The first and second time thresholds can be set according to the actual situation, for example, the first time threshold can be set to 4 days and the second time threshold can be set to 8 days. In this case, if three consecutive days of data are missing in a month, data can be filled by interpolation. However, if four or five consecutive days of data are missing, the missing interval is large, so data from the same period last year can be used as a reference for prediction and filling.

[0068] Furthermore, the method also includes: when scheduling the assigned intelligent agents, tracking the execution progress and intermediate results of each subtask in real time, and transmitting the intermediate results and execution progress to the user terminal for real-time preview; and after scheduling is completed, integrating and outputting the execution results of all subtasks.

[0069] The method provided by this invention enables real-time preview and result feedback during task execution. Intermediate results are displayed in real time during task execution, allowing users to understand the task progress promptly. For example, when generating large geological models, users can view each stage of model construction in real time without waiting for the entire model to complete, effectively reducing user waiting time. After scheduling is complete, the execution results of all sub-tasks can be integrated and automatically interpreted to help users understand the meaning of the results. For example, it can automatically interpret the reasons for changes in production data based on large models and provide corresponding suggestions.

[0070] In some embodiments of the present invention, the user is promptly notified when a task is completed or an anomaly occurs, ensuring that the user does not miss the task completion time and can promptly address any problems that arise during task execution. For example, after a daily report is generated, the system automatically sends the report to the relevant personnel's email or mobile device, or sends a notification to the user when anomalies such as missing data, overload, or malfunction occur.

[0071] The intelligent agent scheduling method proposed in this invention achieves intelligent processing throughout the entire process, from user input to task execution and result feedback. Compared with existing intelligent agent scheduling methods, the method of this invention reduces the number of steps required for users to complete the same task, significantly lowering the learning cost and operational complexity for users. For example, when generating oil and gas field development plans, users only need to provide basic requirements, and the system can automatically complete the formulation and optimization of the plan.

[0072] Example 1: This invention utilizes the reasoning capabilities of large models to automatically decompose and orchestrate complex tasks, adapting to the complex business scenarios of the petroleum industry that involve multiple steps and dependencies, such as oil and gas field development plan formulation and reserve assessment. Steps S11 to S15 are the specific implementation steps in the petroleum industry.

[0073] Step S11, Natural Language-Driven Task Decomposition: Automatically decompose complex tasks described by the user in natural language into multiple simple subtasks.

[0074] The user inputs "Formulate an annual development optimization plan for Oilfield I, requiring analysis of production potential, proposal of adjustment measures, and generation of a detailed report based on existing production data, geological models, and market oil prices." The system decomposes this into the following six sub-tasks: (1) Collect production data (production, pressure, water content, etc.) of Oilfield I over the past year; (2) Retrieve existing three-dimensional geological models and reservoir parameters of Oilfield I; (3) Obtain international oil price trend data over the past year; (4) Analyze the production potential of each block based on production data and geological models; (5) Combine oil price and potential analysis to formulate adjustment measures such as well network adjustment and injection-production parameter optimization; (6) Integrate all analysis results and adjustment measures to generate an annual development optimization report.

[0075] Step S12, Intelligent Subtask Sorting: Based on task dependencies and execution efficiency, the execution order of subtasks is intelligently determined.

[0076] In the above development and optimization tasks, subtasks (1) "collecting production data", (2) "retrieving geological models", and (3) "obtaining oil price data" have no dependencies and can be executed in parallel; subtask (4) "production potential analysis" depends on the results of (1) and (2) and needs to be executed after (1) and (2) are completed; subtask (5) "formulating adjustment measures" depends on the results of (3) and (4) and needs to be executed after (3) and (4) are completed; subtask (6) "generating a report" depends on the results of (5) and is the last step to be executed; at the same time, for the parallel subtasks (1), (2) and (3), since the data volume of subtask (2) "retrieving geological models" is large (about 50GB) and takes a long time, the priority of this subtask is increased, and this subtask is started first, and then (1) and (3) are started synchronously.

[0077] Step S13, Multi-agent collaborative work: Coordinate the collaborative work of multiple agents to achieve efficient completion of complex tasks.

[0078] In the above tasks, the “data acquisition agent” is assigned to perform sub-tasks (1), (2) and (3), the “production analysis agent” to perform sub-task (4), the “development scheme optimization agent” to perform sub-task (5), and the “report generation agent” to perform sub-task (6). The system realizes real-time data transmission between the agents through the intermediate data interaction interface. For example, after the “data acquisition agent” completes data collection, it automatically pushes the data to the “production analysis agent” without manual intervention.

[0079] Step S14, Dynamic Task Execution Plan: Dynamically adjust the task execution plan based on the execution status of subtasks and the system status.

[0080] For example, when the "data acquisition agent" was acquiring production data for oilfield I, it discovered that some wells had missing daily production data. The system automatically adjusted the plan and added a "missing data completion" sub-task, which was executed by the data repair agent. At the same time, the execution of the "production analysis agent" was delayed.

[0081] For example, during the execution of the "development scheme optimization agent", if system resources are tight, the system can split the subtask to be executed into two parallel subtasks, "block 1-5 optimization" and "block 6-10 optimization", and assign them to two backup optimization agents to ensure that the task is completed on time.

[0082] Step S15, Automatic Result Integration: The execution results of multiple subtasks are automatically integrated into the final result.

[0083] The "Report Generation Agent" receives adjustment measures from the "Development Plan Optimization Agent" and potential analysis data from the "Production Analysis Agent." It automatically integrates data charts, analysis conclusions, and recommended measures according to the standard format of petroleum industry development plan reports, generating a complete and logically clear annual development optimization report. Furthermore, if data conflicts exist between different sub-tasks (such as discrepancies between reservoir thickness in the geological model and well logging statistics), the system automatically marks the conflict points and, based on the domain knowledge graph, provides conflict cause analysis and correction suggestions, integrating them into the final report.

[0084] This invention also provides a device for agent scheduling, for example, such as... Figure 2 As shown, the device includes a memory and a processor, the processor being used to run a program, wherein the program is executed to perform the aforementioned agent scheduling method.

[0085] The processor contains a kernel, which retrieves the corresponding program units from memory. One or more kernels can be configured, and agent scheduling is achieved by adjusting kernel parameters.

[0086] The memory may include non-permanent memory in computer-readable media, such as random access memory (RAM) and / or non-volatile memory, such as read-only memory (ROM) or flash RAM, and the memory includes at least one memory chip.

[0087] This invention provides a storage medium on which a program is stored, which, when executed by a processor, implements the agent scheduling method.

[0088] This invention provides a processor for running a program, wherein the program executes the agent scheduling method during runtime.

[0089] This application also provides a computer program product that, when executed on a data processing device, is adapted to perform the steps of initializing the agent scheduling method.

[0090] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0091] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0092] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0093] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0094] In a typical configuration, a computing device includes one or more processors (CPU), input / output interfaces, network interfaces, and memory.

[0095] Memory may include non-persistent memory in computer-readable media, such as random access memory (RAM) and / or non-volatile memory, such as read-only memory (ROM) or flash RAM. Memory is an example of computer-readable media.

[0096] Computer-readable media includes both permanent and non-permanent, removable and non-removable media that can store information using any method or technology. Information can be computer-readable instructions, data structures, modules of programs, or other data. Examples of computer storage media include, but are not limited to, phase-change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, CD-ROM, digital versatile optical disc (DVD) or other optical storage, magnetic tape, magnetic disk storage or other magnetic storage devices, or any other non-transferable medium that can be used to store information accessible by a computing device. As defined herein, computer-readable media does not include transient computer-readable media, such as modulated data signals and carrier waves.

[0097] It should also be noted that the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes that element.

[0098] The above are merely embodiments of this application and are not intended to limit the scope of this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the scope of the claims of this application.

Claims

1. A method for scheduling intelligent agents, characterized in that, The method includes: Obtain user input, the type of which includes text, voice, and / or images; The input is preprocessed to obtain model input information; Based on the large language model, the input information of the model is analyzed to obtain task instructions, which include the task type, parameter information and dependency relationship of at least one subtask. Based on the dependencies, the execution order of the subtasks is determined; Assign an agent to each subtask based on its task type; and Based on the execution order and parameter information, the assigned agents are scheduled.

2. The method according to claim 1, characterized in that, The input is preprocessed to obtain model input information including: Based on the encoder, each type of input is converted into a feature vector, the encoder including a text encoder, a speech encoder and / or an image encoder; If the input type is a single type, then the transformed feature vector is used as the model input information; and If the type of the input is greater than or equal to two, the multiple feature vectors obtained by the conversion will be fused across modalities, and the fused multimodal feature vectors will be used as the input information of the model.

3. The method according to claim 1, characterized in that, Based on the large language model, the input information of the model is analyzed to obtain task instructions, including: The model input information is subjected to intent recognition to obtain the task intent; The task intent is decomposed to obtain at least one sub-task and its task type, wherein the task type includes data statistics, data fusion, report generation, data analysis, 3D modeling, numerical simulation and / or visualization. Extract parameter information for each subtask from the model input information. The parameter information includes oil and gas reservoir name, block range, time period, data type, output format and / or industry standard requirements. Based on a pre-built knowledge graph, the parameter information of each sub-task is completed; Analyze each subtask after completing the parameter information, and determine the input and output data of each subtask; and Based on the input and output data, the subtasks are analyzed to obtain the dependencies between them.

4. The method according to claim 3, characterized in that, Intent recognition of the model input information also includes: The task intent is detected to determine if there are any abnormal issues, including ambiguity, conflict, or error; and If an exception exists, exception handling is performed until the task intent is free of exceptions. The exception handling includes: Display error messages to the user; Receive new input from the user; Based on context awareness and the user's new input, intent recognition is performed to obtain a new task intent.

5. The method according to claim 1, characterized in that, The task instruction also includes the priority of subtasks, and the method further includes: After determining the execution order of the subtasks, it is determined whether there are subtasks that can be executed in parallel; and If they exist, the parallel-executed subtasks are sorted according to their priority to obtain a new execution order. The priorities include low priority, medium priority, and high priority.

6. The method according to claim 1, characterized in that, Based on the task type of each subtask, an agent is assigned to each subtask, including: Perform the following processing on each of the subtasks: Based on the task type of the subtask, at least one candidate agent is matched; Obtain the state information of each candidate agent, including current load, historical execution data, resource usage, health status and / or functional adaptability; Based on the state information, a multi-dimensional score is calculated for each candidate agent; and The candidate agent with the highest multi-dimensional score is selected as the optimal agent, and the optimal agent is assigned to the subtask.

7. The method according to claim 6, characterized in that, Scheduling the assigned agents based on the execution order and parameter information includes: Based on the execution order and parameter information, the intelligent agent corresponding to each subtask is started to execute the subtask, and the resource utilization and fault problems of each intelligent agent are monitored in real time during the execution process. The resource utilization includes the CPU utilization, memory occupancy and / or storage occupancy of the intelligent agent on the system. Based on the resource utilization, determine whether the intelligent agent is overloaded; If any of the aforementioned agents is overloaded, it is checked whether the agent has any subtasks to be executed. If so, the subtasks to be executed are transferred to backup agents, which include the agents among the candidate agents; and If any of the aforementioned agents fails, all of that agent's subtasks are transferred to the backup agent.

8. The method according to claim 7, characterized in that, The method further includes: Collect information on resource utilization and failures of each agent during execution; and The historical execution data is updated based on the collected resource utilization and fault issues.

9. The method according to claim 7, characterized in that, When the agent is a data acquisition agent, the subtasks include: Based on the parameter information, connect to the data source and collect the required data, wherein the data source includes an oilfield SCADA system or a database; The collected data is preprocessed, including data cleaning and data verification. Submit the preprocessed data to the preset path for saving; The integrity of the preprocessed data is checked. If missing data is found, a new missing data completion subtask is added. The missing data completion subtask contains detailed information about the missing data, including hash number, time range and / or missing type. Set the missing data completion subtask to high priority and assign it to the data repair agent for execution; and Identify subtasks that depend on the data acquisition agent, pause their execution until the data repair agent completes its execution.

10. The method according to claim 9, characterized in that, The data repair agent performs the missing data completion subtask, including: Based on the detailed information of the missing data, determine the time when the missing data was missing; Based on the missing time, the missing data is filled in to obtain the repaired data; The rationality of the repair data is verified based on historical data and / or adjacent well data; and If the rationality verification passes, a repair report is generated, and the repair data and the repair report are submitted to a preset path for saving. The repair report includes a summary of the repaired data and the results of its rationality verification.

11. The method according to claim 10, characterized in that, Based on the missing time, completing the missing data includes: If the missing time is less than the first time threshold, then linear interpolation is used to complete the missing data; If the missing time is greater than or equal to a first time threshold and less than a second time threshold, then the missing data is predicted and completed by combining historical data from the same period and data from adjacent wells; and If the missing time is greater than or equal to the second time threshold, then the missing data is predicted and filled in based on the pre-trained prediction model.

12. The method according to claim 1, characterized in that, The method further includes: When scheduling the assigned intelligent agents, the execution progress and intermediate results of each subtask are tracked in real time, and the intermediate results and execution progress are transmitted to the user terminal for real-time preview; and After scheduling is complete, the execution results of all subtasks are integrated and output.

13. A device for agent scheduling, the device comprising a memory and a processor, characterized in that, The processor is used to run a program, wherein the program is run to perform: the method as described in any one of claims 1-12.

Images