An intelligent agent-based autonomous production method for pumping unit wells

Abstract

The application discloses an intelligent agent-based autonomous production method for pumping unit wells, wherein the intelligent agent is composed of five modules, namely, a perception module, a core module, a memory module, an execution module and an interaction module; the perception module collects dynamometer card, pickup, camera, electric parameter and pressure sensor data to realize real-time perception of the state of the pumping unit well; the core module takes the large model reasoning and judgment capability as the kernel, combines the body history production data, fault diagnosis experience and production optimization knowledge stored in the memory module, realizes production operation state monitoring and production energy efficiency optimization of the pumping unit well, the execution module undertakes the execution of the production optimization strategy and the emergency disposal scheme, and the interaction module supports voice and text modes to communicate with the outside world. The intelligent agent-based autonomous production method for pumping unit wells realizes real-time monitoring of the production state and optimization of the production efficiency throughout the life, and provides a new technical means for intelligent, fine and low-carbon production of the pumping unit well.

Description

Technical Field

[0001] This invention relates to the field of oilfield pumping well production management. Specifically, it introduces emerging intelligent agent technology from the field of artificial intelligence to endow pumping wells with the ability to autonomously perceive, make decisions, and execute. Through the reasoning and judgment of large models and the ability to call tools, it completes the planning and execution of tasks, realizes real-time monitoring of the pumping well's operating status and dynamic optimization of production efficiency, and ultimately achieves autonomous management of pumping wells to realize an unmanned management mode. Background Technology

[0002] As a national strategic resource and closely related to the national economy and people's livelihood, the intelligent and low-carbon exploitation of oil resources has always been a key research area. In my country's onshore oilfields, the main types of pumping units used in the middle and late stages of oil well exploitation are beam pumping units and tower pumping units. The daily maintenance and production efficiency optimization of these two types of pumping unit wells are one of the core tasks of oilfield workers. Due to the large number and scattered distribution of pumping unit wells, relying on manual daily maintenance will have the following problems: (1) high labor costs; (2) inability to achieve 24-hour operation and maintenance, and untimely fault repair; (3) fault diagnosis and production optimization require professional technicians with high knowledge thresholds; (4) safety risks exist.

[0003] Research has found that the development of IoT and AI technologies has alleviated some of the problems encountered in manual operation and maintenance. The construction of 5G and fiber optic signals near wells enables remote viewing and transmission of sensor data, microphone data, and monitoring data installed on pumping units. AI algorithms deployed at the edge and server ends can continuously analyze pumping unit well data. Even high-tech solutions such as drones and robotic dogs are used in the daily maintenance of pumping unit wells. All of these technologies can serve as supplementary means to manual management of pumping unit wells. However, there are still areas for improvement. For example, remotely viewing cameras requires inspectors to check the monitoring screens sequentially, which is still a significant workload. Using drones for inspection is costly, difficult to operate, and can only cover pumping units within a certain distance. The results of AI algorithm analysis require further review and processing by professionals, posing a high barrier to entry for ordinary oilfield workers. Therefore, the approach of primarily manual management supplemented by intelligent algorithms still has considerable room for improvement.

[0004] The leap in large-scale model capabilities has spurred the rise of intelligent agent technology. By introducing intelligent agent technology into the management of pumping wells, the current passive management is transformed into autonomous management, providing a new approach to solving current pain points. Specifically, pumping wells are equipped with dedicated intelligent agents. These agents, with large-scale model task planning and reasoning capabilities at their core, autonomously utilize toolsets and professional knowledge bases based on daily operation and maintenance and production efficiency optimization requirements. They perform self-health checks and production optimization, promptly report faults, and implement optimization plans under local control, ultimately achieving safe, low-carbon, and intelligent autonomous production. Summary of the Invention

[0005] To address the aforementioned problems, this invention discloses an autonomous production method for pumping wells based on intelligent agents. By introducing intelligent agent technology into the management of pumping wells, the current passive management is transformed into autonomous management, providing a new technical means for intelligent, refined, and low-carbon production of pumping wells. The intelligent agent consists of five modules: perception, central processing, memory, execution, and interaction. The method can be described as follows:

[0006] An autonomous production method for pumping wells based on intelligent agents, wherein the intelligent agent includes a perception module, a central module, a memory module, an execution module, and an interaction module, wherein:

[0007] The sensing module has the function of multi-dimensional data acquisition from pumping wells and supports a standard interface for calling large model tools;

[0008] The central module, acting as the brain of the intelligent agent, possesses the capabilities of task planning, intelligent analysis, tool invocation, and knowledge retrieval, and can autonomously realize the production management of oil pumping wells and execute user interaction commands.

[0009] The memory module has the function of storing historical production data of pumping wells, fault diagnosis experience, and production optimization knowledge, and supports a standard interface for large model knowledge retrieval.

[0010] The execution module has the ability to control the location of the pumping unit well and supports standard interfaces that conform to the calling of large model tools;

[0011] The interaction module serves as the window for the agent to communicate with the outside world, and is used to receive user commands and return the results of the agent's operation.

[0012] Preferably, the intelligent agent uses the MCP protocol as the standard interface for calling large model tools, RAGFlow as the knowledge base management framework, and the large model uses the DeepSeekV3 model deployed locally on the oilfield computing center platform.

[0013] Preferably, the multi-dimensional data acquisition function of the sensing module includes:

[0014] The raw video stream from the camera next to the pumping unit well is extracted using the RTSP protocol to obtain decoded video frames. ;

[0015] Interact with customized microphone protocols to acquire raw audio PCM data from key locations in pumping wells. ;

[0016] The RTU next to the pumping unit is used to collect dynamometer diagrams, electrical parameters, oil pressure, casing pressure, back pressure, stroke count, and stroke data.

[0017] Preferably, the service interfaces of the perception module include GetCameraFrame for camera video stream capture, GetDataFromRTU for RTU data acquisition, and GetAudioData for audio data acquisition. All interfaces supported by the service are described in tools, and the input parameters, usage methods, and return parameters of the interfaces are annotated in description to help the MCP client know how to use them.

[0018] Preferably, the service interfaces of the memory module include GetOilData (pump well data), GetDiagKnowledge (fault diagnosis expert experience), and GetOptKnowledge (production operation optimization expert knowledge). All interfaces supported by the service are described in tools, and the input parameters, usage methods, and return parameters of the interfaces are annotated in description to help the MCP client know how to use them. The data vectorization storage of the memory module is implemented using the RAGFlow framework.

[0019] Preferably, the central module uses DeepSeekV3 as the brain of the intelligent agent, and has built-in MCP services that implement the algorithm analysis AnalysisServer, including DiagramAnalysis for dynamometer diagram recognition, FluidCalculate for oil measurement, VisionRecognition for key points of the pumping unit, and AudioRecognition for audio anomaly recognition. The service starts and creates MCP clients SenseDataClient for the perception module, StorageDataClient for the memory module, ExecuteClient for the execution module, and AnalysisClient for the algorithm analysis module, as tools that can expand the large model.

[0020] Preferably, the central module is specifically described as follows:

[0021] The detection and perception module, execution module, and memory module support toolsets and knowledge sets.

[0022] It accepts input instructions from the interactive module and performs task planning based on semantic analysis;

[0023] Customized autonomous execution skills This includes, but is not limited to, operational status inspection, production efficiency optimization, and liquid production stability analysis.

[0024] The results of task execution are fed back to the interactive module, including but not limited to fault reporting, fluctuation analysis reporting, knowledge feedback, and suggestion feedback;

[0025] Built-in intelligent analysis set that conforms to the calling interface of large models Including but not limited to indicator diagram recognition Calculation of liquid production Electrophysiological diagnostics Sensor data analysis Identification of key locations of oil pumping units Audio anomaly detection The large model automatically plans and calls specific tools to complete tasks based on the content described by the intelligent analysis tools.

[0026] Preferred, The skill integrates the daily operation and maintenance inspection and energy efficiency optimization of oil pumping wells into the daily fixed work content of the intelligent agent. At the same time, it accepts user instructions and supports the completion of temporary analysis tasks by responding to user input. Both modes are driven by text input to drive the intelligent agent to autonomously complete the response content.

[0027] Preferably, the service interfaces of the execution module include StopOil, OpenOil, and JugeFrequence. All interfaces supported by the service are described in tools, and the input parameters, usage methods, and return parameters of the interfaces are annotated in description to help the MCP client know how to use them.

[0028] Preferably, the interaction module supports both voice and text input. Voice input is converted into text using an ASR tool, and text feedback and TTS playback are also supported. The ASR and TTS functions are implemented using the FunASR open-source project.

[0029] Beneficial effects

[0030] This invention discloses an autonomous production method for pumping wells based on intelligent agents. By introducing intelligent agent technology into the management of pumping wells, the current passive management is transformed into autonomous management, realizing the monitoring of the production operation status and optimization of production energy efficiency of pumping wells. This can effectively shorten the timeliness of fault alarms, reduce the intensity of oilfield workers, and provide a new technical means for intelligent, refined, and low-carbon production of pumping wells. Attached Figure Description

[0031] Figure 1 A schematic diagram of the module structure of an agent-based autonomous production method for oil pumping wells.

[0032] Figure 2 Network configuration diagram for agent operation.

[0033] Figure 3 Core architecture diagram of an intelligent agent.

[0034] Figure 4 Design diagram of the perception module implementation.

[0035] Figure 5 Implementation design diagram of the memory module.

[0036] Figure 6 Design diagram of the implementation of the central module.

[0037] Figure 7 Implementation design diagram of the execution module.

[0038] Figure 8 Flowchart of the execution process of an intelligent agent.

[0039] Specific implementation process

[0040] The following description, in conjunction with the accompanying drawings, illustrates the implementation effects of the agent-based autonomous production method for pumping wells proposed in this invention. It should be understood that these embodiments are for illustrative purposes only and are not intended to limit the scope of the invention. After reading this invention, any modifications of the invention in various equivalent forms by those skilled in the art fall within the scope defined by the appended claims.

[0041] Based on the daily operation and maintenance and production optimization of beam pumping wells, an intelligent agent is constructed according to the method proposed in this invention. The intelligent agent runs on an edge computing terminal next to the pumping well. Figure 1 The core components of the intelligent agent are presented. Python was used for agent development, the MCP protocol was chosen as the standard interface for calling large model tools, RAGFlow was used as the knowledge base management framework, and the large model used was the DeepSeekV3 model deployed locally on the oilfield computing center platform. The network configuration for agent operation is as follows: Figure 2 As shown, the core architecture diagram of the intelligent agent illustrates key technologies, storage, and protocol technologies, such as... Figure 3As shown, the implementation of each module involved in the invention and the operation mode of the intelligent agent are as follows: Speech-to-Text ASR, Text-to-Speech TTS, the open-source MCP architecture FastMCP, the RAG enhancement framework RAGFlo, the vector database Milvus, the temporal database InfluxDB, the relational database MySQL, the video streaming protocol RTSP, the information exchange protocol MQTT, the network protocol HTTP, and the data acquisition protocol ModbusTCP.

[0042] (1) The perception module is encapsulated into an MCP service named SenseDataServer using FastMCP. The service interfaces include GetCameraFrame for camera video stream capture, GetDataFromRTU for RTU data acquisition, and GetAudioData for audio data acquisition. All interfaces supported by the service are described in tools. Detailed descriptions of the input parameters, usage methods, and return parameters of the interfaces are provided in the description to help the MCP client understand how to use them. The implementation design diagram of the perception module is shown below. Figure 4 As shown; the sensing module has the following description:

[0043] Supports extracting raw video streams from cameras near pumping unit wells and obtaining decoded video frames via the RTSP protocol. ;

[0044] Supports integration with customized microphone protocols to acquire raw audio PCM data from key locations in pumping wells. ;

[0045] Supports docking with the RTU next to the pumping unit to collect dynamometer diagrams, electrical parameters, oil pressure, casing pressure, back pressure, stroke count, and stroke data;

[0046] The standard interface supports large model tool calls, enabling the large model to detect all data acquisition toolkits supported by the perception module. , Video tools Audio representation tools, This refers to sensor data tools. The large model autonomously decides to call a specific data acquisition function based on the detected tool's function description.

[0047] (2) The memory module is encapsulated into an MCP service named StorageDataServer using FastMCP. The service interfaces include GetOilData (pump well data), GetDiagKnowledge (expert experience in fault diagnosis), and GetOptKnowledge (expert knowledge in production operation optimization). All interfaces supported by the service are described in tools. The input parameters, usage methods, return parameters, etc. of the interfaces are explained in the description, so that the MCP client knows how to use them. The data vectorization storage of the memory module is implemented using the RAGFlow framework. The implementation design diagram of the memory module is shown below. Figure 5 As shown; the memory module is specifically described as follows:

[0048] It supports the storage of pumping unit well production data, including well parameters, dynamometer diagrams, electrical parameters, pressure, fluid production, production execution strategies, and failure cases;

[0049] Supports expert experience in diagnosing pumping unit well faults, including rod breakage, tubing leakage, wax buildup, pump jamming, belt breakage, and heavy oil.

[0050] Supports storing expert knowledge for optimizing the production and operation of pumping wells, including stroke optimization, intermittent pumping optimization, balancing, and peak-shifting optimization;

[0051] It supports standard interfaces for knowledge retrieval in large models, and the large model can detect the knowledge sets that the memory module supports for retrieval. , This represents the production data of the pumping unit well. This indicates experience in fault diagnosis. This represents the knowledge module that optimizes the execution of queries, allowing the large model to autonomously plan queries based on the content of the dataset.

[0052] (3) The central module utilizes DeepSeekV3 as the brain of the intelligent agent, and has built-in MCP services for algorithm analysis (AnalysisServer), including DiagramAnalysis for dynamometer diagram recognition, FluidCalculate for oil measurement, VisionRecognition for key points of the pumping unit, and AudioRecognition for audio anomaly recognition. The service startup creates MCP clients for the perception module (SenseDataClient), the memory module (StorageDataClient), the execution module (ExecuteClient), and the algorithm analysis module (AnalysisClient), serving as tools to expand the large model. The implementation design diagram of the central module is shown below. Figure 6 As shown; the central module is specifically described as follows:

[0053] Supports the toolset and knowledge set capabilities supported by the detection and perception module, execution module, and memory module;

[0054] It supports accepting command input from interactive modules and performing task planning based on semantic analysis;

[0055] Supports customized autonomous execution skills This includes, but is not limited to, operational status inspection, production efficiency optimization, and liquid production stability analysis.

[0056] It supports feeding back the results of task execution to the interactive module, including but not limited to fault reporting, fluctuation analysis reporting, knowledge feedback, and suggestion feedback;

[0057] Built-in intelligent analysis set that conforms to the calling interface of large models Including but not limited to indicator diagram recognition Calculation of liquid production Electrophysiological diagnostics Sensor data analysis Identification of key locations of oil pumping units Audio anomaly detection The large model automatically plans and calls upon specific tools to complete tasks based on the content described by the intelligent analysis tools. These built-in analysis algorithms can all be implemented through traditional algorithms or deep learning model training algorithms.

[0058] Dynamogram Recognition By collecting common working conditions of oil wells, the ResNet deep learning algorithm is used for classification training;

[0059] Liquid production calculation The effective stroke point is calculated using the dynamometer diagram based on the 5-point curvature method or deep learning regression algorithm, and the production volume is calculated using the effective stroke method.

[0060] Electrophysiological Diagnosis Sensor data analysis Time series data diagnostics are performed using the data analysis algorithm library SciPy;

[0061] Key position identification of oil pumping unit The YOLO deep learning target detection algorithm is used to detect key locations in oil pumping wells.

[0062] Audio anomaly detection We trained the system to classify sound faults using the open-source algorithm FunASR framework.

[0063] (4) The execution module utilizes FastMCP to encapsulate an MCP service named ExecuteServer. The service interfaces include StopOil, OpenOil, and JugeFrequence. All supported interfaces are described in tools, and detailed information on input parameters, usage, and return parameters is provided in the description to help the MCP client understand how to use them. The implementation design diagram of the execution module is shown below. Figure 7 As shown; the execution module is specifically described as follows:

[0064] Supports well shutdown function;

[0065] Supports stroke rate adjustment function;

[0066] Supports well opening function;

[0067] It supports standard interfaces that conform to the large model tool call, and the large model can detect all control instruction sets supported by the execution module. , This indicates the interface for stopping the well. This indicates the well-opening command interface. This indicates that the specified interface is used for stroke adjustment. The large model can autonomously decide which control function to call based on the description in the interface.

[0068] (5) The interaction module supports both voice and text input. Voice input is converted into text using the ASR tool. It also supports text feedback and TTS broadcast. The FunASR open source project is used to implement ASR and TTS functions.

[0069] (6) The intelligent agent's operation mode is designed to support custom SKILLS skills and accept user instructions to perform work. The custom SKILLS skills will integrate the daily operation and maintenance inspection and energy efficiency optimization of the pumping well into the daily fixed work content of the intelligent agent. Accepting user instructions supports completing temporary analysis tasks by responding to user input. Both modes are driven by text input to drive the intelligent agent to autonomously complete the response content. Figure 8 The operation process of the intelligent agent is demonstrated using the "daily inspection of the pumping unit" command. The daily inspection of the pumping unit checks whether the pumping unit is operating normally, whether the downhole pump is working properly as reflected in the dynamometer, whether there are fluctuations in the production rate, whether there is oil leakage at the packing at the wellhead, and whether the belt is slipping.

[0070] As can be seen from the above analysis, the intelligent agent-based autonomous production method for oil pumping wells proposed in this invention can realize the monitoring of the production operation status and optimization of production energy efficiency of oil pumping wells, providing a new technical means for intelligent, refined and low-carbon production of oil pumping wells.

Claims

1. A method for autonomous production of oil pumping wells based on intelligent agents, characterized in that, An intelligent agent comprises a perception module, a central module, a memory module, an execution module, and an interaction module, wherein: The sensing module has the function of multi-dimensional data acquisition from pumping wells and supports a standard interface for calling large model tools; The central module, acting as the brain of the intelligent agent, possesses the capabilities of task planning, intelligent analysis, tool invocation, and knowledge retrieval, and can autonomously realize the production management of oil pumping wells and execute user interaction commands. The memory module has the function of storing historical production data of pumping wells, fault diagnosis experience, and production optimization knowledge, and supports a standard interface for large model knowledge retrieval. The execution module has the ability to control the location of the pumping unit well and supports standard interfaces that conform to the calling of large model tools; The interaction module serves as the window for the agent to communicate with the outside world, and is used to receive user commands and return the results of the agent's operation.

2. The method according to claim 1, characterized in that, The agent uses the MCP protocol as the standard interface for calling large model tools, RAGFlow as the knowledge base management framework, and the large model uses the DeepSeekV3 model deployed locally on the oilfield computing center platform.

3. The method according to claim 1, characterized in that, The multi-dimensional data acquisition function of the sensing module includes: The raw video stream from the camera next to the pumping unit well is extracted using the RTSP protocol to obtain decoded video frames. ; Interact with customized microphone protocols to acquire raw audio PCM data from key locations in pumping wells. ; The RTU next to the pumping unit is used to collect dynamometer diagrams, electrical parameters, oil pressure, casing pressure, back pressure, stroke count, and stroke data.

4. The method according to claim 1, characterized in that, The service interfaces of the perception module include GetCameraFrame for camera video stream capture, GetDataFromRTU for RTU data acquisition, and GetAudioData for audio data acquisition. All interfaces supported by the service are described in the tools section, and the input parameters, usage methods, and return parameters of the interfaces are annotated in the description section to help the MCP client know how to use them.

5. The method according to claim 1, characterized in that, The service interfaces of the memory module include GetOilData (pump well data), GetDiagKnowledge (expert experience in fault diagnosis), and GetOptKnowledge (expert knowledge in production operation optimization). All interfaces supported by the service are described in tools. The input parameters, usage methods, and return parameters of the interfaces are annotated in description to help the MCP client know how to use them. The data vectorization storage of the memory module is implemented using the RAGFlow framework.

6. The method according to claim 1, characterized in that, The central module uses DeepSeekV3 as the brain of the intelligent agent, and has built-in MCP services that implement the algorithm analysis AnalysisServer, including DiagramAnalysis for dynamometer diagram recognition, FluidCalculate for oil measurement, VisionRecognition for key points of the pumping unit, and AudioRecognition for audio anomaly recognition. The service starts and creates MCP clients SenseDataClient for the perception module, StorageDataClient for the memory module, ExecuteClient for the execution module, and AnalysisClient for the algorithm analysis module, as tools to expand the large model.

7. The method according to claim 1, characterized in that, The central module is specifically described as follows: The detection and perception module, execution module, and memory module support toolsets and knowledge sets. It accepts input instructions from the interactive module and performs task planning based on semantic analysis; Customized autonomous execution skills This includes, but is not limited to, operational status inspection, production efficiency optimization, and liquid production stability analysis. The results of task execution are fed back to the interactive module, including but not limited to fault reporting, fluctuation analysis reporting, knowledge feedback, and suggestion feedback; Built-in intelligent analysis set that conforms to the calling interface of large models Including but not limited to indicator diagram recognition Calculation of liquid production Electrophysiological diagnostics Sensor data analysis Identification of key locations of oil pumping units Audio anomaly detection The large model automatically plans and calls specific tools to complete tasks based on the content described by the intelligent analysis tools.

8. The method according to claim 7, characterized in that, The skill integrates the daily operation and maintenance inspection and energy efficiency optimization of oil pumping wells into the daily fixed work content of the intelligent agent. At the same time, it accepts user instructions and supports the completion of temporary analysis tasks by responding to user input. Both modes are driven by text input to drive the intelligent agent to autonomously complete the response content.

9. The method according to claim 1, characterized in that, The service interfaces of the execution module include StopOil, OpenOil, and JugeFrequence. All interfaces supported by the service are described in tools, and the input parameters, usage methods, and return parameters of the interfaces are commented in description to help the MCP client know how to use them.

10. The method according to claim 1, characterized in that, The interactive module supports both voice and text input. Voice input is converted to text using an ASR tool, and it also supports text feedback and TTS playback. The ASR and TTS functions are implemented using the FunASR open-source project.

Images