Pumping unit lubrication state self-learning system and method based on multi-mode sensing

Abstract

The invention discloses a pumping unit lubrication state self-learning system and method based on multi-modal sensing, the system synchronously collects data such as pumping unit vibration spectrum, oil metal particle concentration, temperature, pressure and the like through multi-modal sensing, and after data adaptive transmission, a deep reinforcement learning model is combined with time sequence prediction and an oil film thickness algorithm to obtain the lubrication state of a pumping unit. Deeply analyzing the data to predict a lubrication state trend; the self-adaptive dynamic oil supply algorithm dynamically adjusts the grease injection pressure and the oil quantity according to the working condition by combining pulse type high-pressure grease injection and fuzzy PID control logic according to a model decision. The system constructs a closed loop from data acquisition, intelligent analysis to accurate execution, guarantees data security by using block chain evidence storage, and realizes policy sharing between devices based on the cloud. The problems that traditional lubrication management data collection is single and lags behind, and an oil supply strategy lacks dynamic adaptability are solved, intelligent and accurate management of the lubrication state of the oil pumping unit is achieved, and the equipment fault and maintenance cost is reduced.

Description

Technical Field

[0001] The present invention relates to the field of oil pumping unit lubrication, and in particular to an oil pumping unit lubrication state self-learning system and method based on multi-modal sensing. Background Art

[0002] In oilfield production, pumping units operate in complex and ever-changing environments. Accurate management of lubrication status is crucial for ensuring stable equipment operation, extending service life, and reducing maintenance costs. With the increasing trend toward intelligent and automated oilfield operations, traditional management methods relying on manual inspections and regular grease refilling are no longer adequate for modern, efficient production. Advanced technologies are urgently needed to achieve intelligent lubrication status management.

[0003] Existing oil pump lubrication management technology has many shortcomings. On the one hand, the data collection method is single and lagging. Traditional methods mostly rely on manual experience and judgment or a small number of single-point sensors. It is difficult to fully capture multi-dimensional information such as vibration, temperature, and oil composition during the operation of the oil pump. It is unable to perceive subtle changes in the lubrication status in time, resulting in the delay in the discovery of abnormal lubrication problems, which can easily lead to equipment failure. On the other hand, the oil supply strategy lacks dynamic adaptability. The traditional timed and quantitative oil supply mode does not fully consider the differences in the actual operating conditions of the oil pump. Under conditions of load changes, speed fluctuations, etc., it is easy for excessive oil supply to cause waste and pollution, or insufficient oil supply to aggravate equipment wear, and it is impossible to achieve the optimal allocation of lubrication resources. Summary of the Invention

[0004] In order to overcome the shortcomings and deficiencies of the prior art, the present invention provides a self-learning system and method for the lubrication status of an oil pump based on multimodal sensing.

[0005] The technical solution adopted by the present invention is a self-learning system for the lubrication status of an oil pumping unit based on multimodal sensing, comprising:

[0006] Multimodal composite parameter acquisition unit, which includes a high-precision vibration spectrum analysis module, a nano-scale oil and metal particle detection module, and a temperature-pressure dual-mode sensing node. This unit adopts a double-layer sealed cavity design;

[0007] The data transmission and preprocessing adaptation unit is used to encode and format-convert the data collected by the multimodal composite parameter acquisition unit, and transmit the data to subsequent units through a preset communication protocol;

[0008] The deep reinforcement learning model construction unit integrates the LSTM time series prediction module to learn the time series characteristics of the collected data. At the same time, it combines the oil film thickness prediction algorithm to build a model architecture based on deep reinforcement learning.

[0009] The adaptive dynamic oil supply algorithm execution unit is equipped with pulsed high-pressure grease injection technology and combines fuzzy PID control logic to adjust the oil injection volume. This unit interacts with the deep reinforcement learning model construction unit and receives the oil supply strategy instructions output by the model;

[0010] The decision-making result output and execution drive unit converts the lubrication parameter optimization decision generated by the deep reinforcement learning model building unit and the oil supply strategy determined by the adaptive dynamic oil supply algorithm execution unit into a control signal that drives the actuator of the oil pumping unit's lubrication system;

[0011] The blockchain data evidence management unit is used to record the lubrication operation logs throughout the entire life cycle of the pumping unit and securely store and share data based on blockchain technology. This unit interacts with other units in the system to obtain relevant lubrication operation information;

[0012] The cloud-based collaborative communication interaction unit, based on the 5G-MEC edge computing node, conducts policy sharing communication between device groups. This unit interacts with the deep reinforcement learning model construction unit and the blockchain data storage management unit to complete cloud-based data transmission and reception.

[0013] Furthermore, the deep reinforcement learning model construction unit constructs a deep reinforcement learning model that uses the following formula to predict the oil film thickness:

[0014]

[0015] Among them, H t+1 represents the predicted oil film thickness at time t+1; is the eigenvector of the vibration spectrum of the pumping unit at time t after fast Fourier transform; M t is the concentration vector of metal particles in the oil at time t; T t is the temperature value at time t; P t is the pressure value at time t; α t , β t , γ t , δ t is a time-varying weight coefficient determined by the model through learning; θ is the parameter set of the model; f is a nonlinear mapping function based on deep learning, which realizes the prediction of oil film thickness by fusion processing of multimodal data.

[0016] Furthermore, the deep reinforcement learning model construction unit adopts the following reward function formula when optimizing lubrication parameters:

[0017]

[0018] Among them, R t represents the reward value at time t; Et is the energy consumption value of the pumping unit at time t; W t is the wear rate of key components of the pumping unit at time t; L t is the oil injection amount at time t; ω1, ω2, and ω3 are weight coefficients, and satisfy ω1+ω2+ω3=1. The weight coefficients are adjusted to achieve an optimal balance among the three objectives of energy consumption, wear rate, and oil injection amount.

[0019] Furthermore, the adaptive dynamic oil supply algorithm execution unit has a pulsed high-pressure grease injection pressure adjustment formula as follows:

[0020] P inj =P0+k1·ΔV+k2·ΔT+k3·ΔM

[0021] Among them, P inj is the actual output grease injection pressure; P0 is the initial set pressure; ΔV is the difference in vibration spectrum characteristics between the current moment and the previous moment; ΔT is the temperature difference between the current moment and the previous moment; ΔM is the change in metal particle concentration in the oil between the current moment and the previous moment; k1, k2,

[0022] k3 is the pressure regulation coefficient, which is determined through experiments and model learning according to different operating conditions and equipment parameters of the pumping unit.

[0023] Furthermore, the adaptive dynamic oil supply algorithm execution unit, the oil injection amount adjustment formula is based on fuzzy PID control, as shown below:

[0024]

[0025] Wherein, ΔL is the adjustment amount of oil injection; e is the deviation between the current oil injection amount and the target oil injection amount; K p is the proportionality coefficient, K i is the integral coefficient, K d are differential coefficients, which are dynamically adjusted according to the load and operating time parameters of the pumping unit through fuzzy logic rules.

[0026] Furthermore, the deep reinforcement learning model construction unit and the adaptive dynamic fuel supply algorithm execution unit work together through the following information interaction mechanism:

[0027] A t =π θ (S t )

[0028] Among them, A t is the refueling action strategy output by the deep reinforcement learning model at time t; S t is the state vector collected by the multimodal composite parameter acquisition unit at time t and input into the deep reinforcement learning model after being processed by the data transmission and preprocessing adaptation unit;θ It is a policy network based on the parameters θ of the deep reinforcement learning model. The oil supply action strategy is generated through the network, and then the adaptive dynamic oil supply algorithm execution unit converts the strategy into a specific grease injection operation.

[0029] Furthermore, the data processing relationship formula between the multimodal composite parameter acquisition unit and the deep reinforcement learning model construction unit is:

[0030]

[0031] in, is the state vector input to the deep reinforcement learning model; is the original data vector collected by the multimodal composite parameter acquisition unit; φ is the data processing parameter set; g is the data processing function, which converts the original data into a state vector form suitable for deep reinforcement learning model input by performing feature extraction and normalization operations on the original data.

[0032] Furthermore, when the adaptive dynamic oil supply algorithm execution unit takes into account the influence of the pumping unit's operating speed, the oil injection amount correction formula is:

[0033] L corr =L base (1+k v v)

[0034] Among them, L corr is the corrected oil filling amount; L base is the basic oil injection volume; v is the operating speed of the pumping unit; k v is the speed influence coefficient, which is determined through experiments and model training according to the mechanical structure and lubrication requirements of the pumping unit.

[0035] Furthermore, the deep reinforcement learning model construction unit adopts the following state transition probability formula when combining different working stages of the pumping unit:

[0036]

[0037] Among them, P(S t+1 ∣S t ,A t ) indicates that the state at time t is S t , perform action A t After that, it transfers to state S at time t+1 t+1 The probability of p i is the probability weight associated with different working stages of the pumping unit; f i (S t ,A t) is the state transfer function corresponding to different working stages, which is obtained by learning the historical data of the pumping unit in different working stages to simulate the system state transfer process.

[0038] The self-learning method of the lubrication state of the oil pumping unit based on multimodal sensing includes the following steps:

[0039] In the first step, a multi-modal composite parameter acquisition unit is used to collect the vibration spectrum, oil metal particles, temperature and pressure parameters during the operation of the pumping unit in real time;

[0040] In the second step, the collected multimodal parameter data is format converted and encoded by the data transmission and preprocessing adaptation unit, and then transmitted to the deep reinforcement learning model construction unit;

[0041] In the third step, the deep reinforcement learning model construction unit uses the LSTM time series prediction module to learn the time series characteristics of the data based on the received data. Combined with the oil film thickness prediction algorithm, it constructs a deep reinforcement learning model to generate lubrication parameter optimization decisions.

[0042] In the fourth step, the adaptive dynamic oil supply algorithm execution unit receives the decision output by the deep reinforcement learning model construction unit and determines the specific oil injection pressure and oil injection volume adjustment strategy based on the pulsed high-pressure oil injection technology and fuzzy PID control logic;

[0043] In the fifth step, the decision result output and execution drive unit converts the optimization decision and oil supply strategy into control signals, driving the actuators of the oil pumping unit lubrication system to perform corresponding actions;

[0044] In the sixth step, the blockchain data evidence management unit records relevant information during the entire lubrication operation and securely stores and shares the data based on blockchain technology.

[0045] In the seventh step, the cloud-based collaborative communication interaction unit conducts strategy sharing and data interaction between the system and other pumping equipment based on the 5G-MEC edge computing node.

[0046] Beneficial effects: The present invention proposes a self-learning system and method for the lubrication status of an oil pump based on multimodal sensing. The system utilizes multimodal sensing technology to synchronously collect multi-dimensional data such as the vibration spectrum of the oil pump, metal particle concentration in the oil, temperature and pressure. Compared with the traditional single data collection method, it can capture changes in the lubrication status more comprehensively and timely, and avoid potential faults caused by information lag. The deep reinforcement learning model combines time series prediction with the oil film thickness algorithm to conduct an in-depth analysis of the collected data, accurately predict the lubrication status trend, and break through the subjectivity and limitations of traditional empirical judgments. In terms of oil supply strategy, the adaptive dynamic oil supply algorithm comprehensively considers the real-time working conditions of the oil pump, and dynamically adjusts the grease injection pressure and oil volume according to changes in parameters such as vibration, temperature, and oil composition, completely changing the drawbacks of the traditional timed and quantitative oil supply mode, avoiding the waste of resources and environmental pollution caused by excessive oil supply, and preventing equipment wear caused by insufficient oil supply, thereby achieving precise allocation of lubrication resources. This invention builds a closed-loop management system from data collection and intelligent analysis to precise execution, combines blockchain technology to ensure data security and reliability, and realizes strategy sharing between devices based on the cloud, which significantly improves the intelligence and refinement of pumping unit lubrication management, effectively reduces equipment failure rate and maintenance costs, and provides strong support for efficient and stable production in oil fields. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] Figure 1 It is a diagram of the system unit composition of the present invention;

[0048] Figure 2 The figure is a flow chart of the method steps of the present invention. DETAILED DESCRIPTION

[0049] It should be noted that, unless there is a conflict, the embodiments and features in the embodiments of this application can be combined with each other. The present application will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0050] like Figure 1 As shown, the oil pumping unit lubrication state self-learning system and method based on multimodal sensing includes:

[0051] The multimodal composite parameter acquisition unit includes a high-precision vibration spectrum analysis module with a frequency response range of 0-5kHz, used to collect the vibration spectrum of the oil pump during operation; a nano-scale oil metal particle detection module with a detection accuracy of 0.5μm, which can detect metal particles in the oil; and a temperature-pressure dual-mode sensing node with a temperature sensitivity of 0.1°C and a pressure sensitivity of 0.01MPa, capable of simultaneously collecting temperature and pressure parameters. The unit adopts a double-layer sealed cavity design to maintain data acquisition integrity in environments with sand content greater than 15%. Each sensing module is connected to the subsequent unit through a preset interface.

[0052] Specifically, the multimodal composite parameter acquisition unit, serving as the system's sensing core, integrates a high-precision vibration spectrum analysis module, a nanoscale oil metal particle detection module, and a temperature-pressure dual-mode sensing node to achieve multi-dimensional monitoring of the pumping unit's lubrication status. The vibration spectrum analysis module's 0-5 kHz frequency response range covers the characteristic vibration frequencies of key components such as the pumping unit's crank, connecting rod, and reduction gearbox under different operating conditions. For example, gearbox gear wear produces characteristic frequencies in the 1-3 kHz range, while bearing failure may produce abnormal peaks in the 3-5 kHz range. The nanoscale oil metal particle detection module's 0.5 μm detection accuracy enables early detection of metal abrasive particles such as copper, iron, and aluminum, providing a basis for assessing the wear status of components such as bearings and gears. A detected iron particle concentration exceeding 50 ppm indicates possible abnormal gear or bearing wear. The temperature-pressure dual-mode sensing node, with its 0.1°C temperature sensitivity and 0.01 MPa pressure sensitivity, enables real-time monitoring of temperature changes and oil supply pressure fluctuations at lubrication points. During the startup phase of the pump, the lubricating oil temperature will rise rapidly. If the temperature rise rate exceeds 0.5℃ / min, it may indicate poor lubrication or excessive load.

[0053] The unit's double-layer sealed chamber design utilizes a stainless steel housing and fluororubber seals, effectively preventing dust from intruding into the sensor even in harsh environments with sand content exceeding 15%, ensuring reliable data acquisition. A nano-scale anti-corrosion coating is applied to the sensor surface, allowing it to withstand the high humidity and salt spray conditions found in oilfields, extending its service life to over five years. Each sensor module connects to the data transmission unit via an RS485 or CAN bus, utilizing the Modbus protocol for data communication at speeds up to 115.2 kbps, ensuring stable real-time data transmission.

[0054] The significance of the multimodal composite parameter acquisition unit lies in its ability to overcome the limitations of traditional single-sensor monitoring and achieve comprehensive awareness of lubrication conditions. By integrating data from multiple sources, including vibration, oil, temperature, and pressure, it enables early detection of potential lubrication failures, such as oil film rupture and increased wear, providing an accurate data foundation for subsequent intelligent decision-making. For example, if abnormal peaks appear in the vibration spectrum, the concentration of iron particles in the oil increases, and the temperature rises, the system can comprehensively identify poor lubrication and provide early warning to prevent equipment failure.

[0055] The data transmission and preprocessing adaptation unit is used to perform adaptation operations such as encoding and format conversion on the data collected by the multimodal composite parameter acquisition unit, and transmit the data to subsequent units through a preset communication protocol;

[0056] Specifically, the data transmission and preprocessing adaptation unit is responsible for format conversion, noise filtering, and protocol adaptation of the raw data obtained by the multimodal composite parameter acquisition unit to meet the input requirements of the subsequent deep reinforcement learning model. The unit first digitizes the collected analog signals and converts analog signals such as vibration, temperature, and pressure into digital signals through a 16-bit ADC converter. The sampling frequency is as high as 10kHz to ensure that the signal is not distorted. In the data preprocessing stage, the wavelet transform algorithm is used to reduce the noise of the vibration signal, remove environmental noise and electrical interference, and improve the signal-to-noise ratio of the signal. For oil metal particle data, the Kalman filter algorithm is used for smoothing to eliminate random errors in the detection process.

[0057] For data transmission, the unit supports multiple communication protocols, including Ethernet, 4G / 5G, and LoRa, allowing users to select the appropriate transmission method based on the site environment. In remote oilfield scenarios, 5G communication technology is used to achieve high-speed data transmission with a transmission delay of less than 10ms, ensuring real-time performance. For remote areas, the LoRa protocol can be switched to, achieving communication distances exceeding 10km, meeting wide-area coverage requirements. Data is encrypted using the AES-256 algorithm during transmission to prevent tampering or theft during transmission and ensure data security.

[0058] The implementation of the data transmission and preprocessing adaptation unit involves both hardware and software. The hardware utilizes an industrial-grade ARM processor with a clock speed of up to 1GHz, providing sufficient computing power to process large amounts of sensor data. The software runs an embedded Linux operating system and includes dedicated data processing middleware to enable data collection, processing, storage, and transmission. This unit addresses the issue of fusion of heterogeneous data from multiple sources, converting raw data collected by different sensors into standardized data in a unified format. This provides high-quality input for deep reinforcement learning models, improving their training effectiveness and prediction accuracy.

[0059] The deep reinforcement learning model construction unit integrates the LSTM time series prediction module, which can learn the time series characteristics of the collected data. At the same time, it combines the oil film thickness prediction algorithm to build a model architecture based on deep reinforcement learning. This unit receives data from the data transmission and preprocessing adaptation unit and processes the data into the model input format;

[0060] Specifically, the deep reinforcement learning model construction unit is the intelligent decision-making core of the system. By integrating the LSTM time series prediction module and the oil film thickness prediction algorithm, it conducts in-depth analysis and learning of the pre-processed multimodal data. The LSTM time series prediction module can capture the time series characteristics of the data and predict the changing trends of parameters such as vibration, temperature, and pressure. For example, by analyzing the temperature data of the past 24 hours, the changes in lubricating oil temperature in the next 4 hours can be predicted, and potential temperature anomalies can be discovered in advance. The oil film thickness prediction algorithm is based on the principles of fluid mechanics and machine learning methods, and comprehensively considers factors such as load, speed, and lubricating oil viscosity to predict the oil film thickness between key friction pairs. When the predicted oil film thickness is less than the critical value (such as 0.1μm), it indicates that the lubrication state is poor and the oil supply strategy needs to be adjusted.

[0061] This unit utilizes a layered architecture. The bottom layer is the data processing layer, responsible for feature extraction and normalization of input data. The middle layer is the model training layer, which uses the TensorFlow framework to build a deep reinforcement learning model, using historical data for offline training and online updates. The top layer is the decision output layer, which generates the optimal lubrication parameter adjustment strategy based on the trained model. During model training, an experience replay mechanism and an ε-greedy strategy are used to improve the model's learning efficiency and stability. By continuously interacting with the environment, the model can automatically adjust its parameters to meet the lubrication requirements of different pumping unit operating conditions.

[0062] The significance of the deep reinforcement learning model building unit lies in enabling intelligent decision-making for lubrication management. Traditional rule-based methods struggle to cope with the complex and ever-changing operating conditions of pumping units. However, the deep reinforcement learning model can autonomously learn to discover the complex relationships hidden in the data and generate an optimal oil supply strategy. For example, during the heavy-load startup phase of the pumping unit, the model automatically increases the oil supply and grease injection pressure to ensure sufficient oil film thickness; during stable operation, it appropriately reduces the oil supply to reduce energy consumption and lubrication costs. This adaptive decision-making mechanism significantly improves lubrication efficiency and equipment reliability.

[0063] The adaptive dynamic oil supply algorithm execution unit is equipped with pulsed high-pressure grease injection technology, with an adjustable pressure range of 0.5-20MPa. Combined with fuzzy PID control logic, it can adjust the oil injection volume. This unit interacts with the deep reinforcement learning model construction unit and receives the oil supply strategy instructions output by the model;

[0064] Specifically, the adaptive dynamic oil supply algorithm execution unit uses pulsed high-pressure grease injection technology and fuzzy PID control logic based on the decision-making results of the deep reinforcement learning model to achieve precise adjustment of the grease injection pressure and oil injection volume. Pulsed high-pressure grease injection technology uses high-frequency solenoid valve control, and the grease injection pressure is adjustable in the range of 0.5-20MPa, which can provide appropriate grease injection pressure according to the needs of different lubrication points. For example, for heavy-loaded components such as reduction gears, the grease injection pressure can be increased to above 15MPa to ensure that the lubricant can penetrate the friction pair surface; for light-loaded components such as bearings, the grease injection pressure can be reduced to 2-5MPa to avoid waste.

[0065] Fuzzy PID control logic dynamically adjusts PID controller parameters by monitoring the oil injection amount deviation and the rate of change of the deviation in real time. During the oil injection process, the system continuously compares the deviation between the actual oil injection amount and the target oil injection amount. When the deviation is large, the proportional coefficient is increased to speed up the response; when the deviation is small, the integral coefficient is increased to eliminate static error. Fuzzy rule tables are used to adjust the PID parameters online, achieving high-precision control of the oil injection amount to ±0.1ml. For example, if the system detects insufficient oil film thickness, it automatically increases the oil injection amount; when the oil film thickness reaches the ideal state, the oil injection amount is maintained stable.

[0066] The implementation of the adaptive dynamic lubrication algorithm execution unit relies on a high-precision grease pump and solenoid valve control system. The grease pump is driven by a servo motor, achieving a flow control accuracy of ±0.5%, capable of meeting lubrication requirements under varying operating conditions. The solenoid valve response time is less than 10ms, ensuring precise control of pulsed grease injection. This unit's significance lies in transcending the limitations of traditional timed and quantitative lubrication, enabling dynamic adjustment of the lubrication strategy based on actual lubrication conditions. By precisely controlling the injection pressure and volume, it ensures adequate lubrication while minimizing lubricant waste, reducing operational costs and extending equipment life.

[0067] The decision-making result output and execution drive unit converts the lubrication parameter optimization decision generated by the deep reinforcement learning model building unit and the oil supply strategy determined by the adaptive dynamic oil supply algorithm execution unit into a control signal that drives the actuator of the oil pumping unit's lubrication system;

[0068] Specifically, the decision output and execution drive unit is responsible for converting the lubrication parameter optimization decisions generated by the deep reinforcement learning model and the oil supply strategy determined by the adaptive dynamic oil supply algorithm into specific control signals to drive the actuators of the oil pump lubrication system. This unit adopts a modular design and includes a digital output module, an analog output module, and a communication interface module. The digital output module provides multiple relay outputs for controlling the start and stop of equipment such as grease pumps and solenoid valves. The analog output module provides a 4-20mA current signal output for adjusting continuous control parameters such as motor speed and pressure regulating valve opening.

[0069] During the signal conversion process, the unit first decodes and analyzes the decision results, converting abstract optimization parameters into concrete physical quantities. For example, the model outputs the oil supply strategy, converting it into parameters such as the grease pump's start and stop times, injection pressure, and injection volume. Signal conditioning and power amplification are then performed based on the actuator's characteristics to ensure the control signal can reliably drive the actuator. For high-power devices, solid-state relays or contactors are used for drive, capable of withstanding voltages of 220V / 380V and currents exceeding 10A.

[0070] The implementation of the decision-making output and execution drive unit involves hardware circuit design and software driver development. An isolation design is employed in the hardware to ensure electrical isolation between control signals and external devices, enhancing the system's anti-interference capabilities. A dedicated driver program was developed in the software to map and convert decision results into control signals. This unit bridges the gap between intelligent decision-making and physical execution, translating abstract optimization strategies into actual equipment actions, ensuring the lubrication system operates according to the optimal solution and ultimately achieving the goal of intelligent lubrication management.

[0071] The blockchain data evidence management unit is used to record the lubrication operation logs throughout the entire life cycle of the pumping unit and securely store and share data based on blockchain technology. This unit interacts with other units in the system to obtain relevant lubrication operation information;

[0072] Specifically, the blockchain data evidence management unit, based on blockchain technology, records and manages lubrication operation logs throughout the pumping unit's lifecycle, ensuring data immutability, traceability, and secure sharing. This unit utilizes a consortium blockchain architecture, with participation from multiple parties, including oilfield management, equipment manufacturers, and operation and maintenance service providers, who jointly maintain the blockchain network. Each node maintains a complete copy of the blockchain, ensuring data consistency through a consensus mechanism. Regarding data recording, the unit packages key information such as the time, injection volume, injection pressure, and equipment status of each lubrication operation into blocks, adds a timestamp, and links them to the blockchain.

[0073] To ensure data security, the blockchain uses asymmetric encryption technology to encrypt data. Each participant possesses a pair of public and private keys: the private key is used to sign data, and the public key is used to verify the validity of the signature. Only authorized users can decrypt and view the relevant data. Furthermore, the distributed nature of the blockchain ensures that the failure or tampering of any single node will not affect the normal operation of the entire system, improving data reliability and security. Regarding data sharing, the unit provides a standardized API interface, through which authorized users can query and obtain historical lubrication data, enabling cross-departmental and cross-enterprise data sharing.

[0074] The implementation of the blockchain data evidence management unit involved the selection of a blockchain platform and the development of smart contracts. The platform chose Hyperledger Fabric as its underlying framework, characterized by high performance, scalability, and security. Smart contracts, developed in the Go language, implement data recording, querying, and permissions management. This unit addresses the challenges of traditional data management methods, such as data tampering and traceability. Using blockchain technology, a trusted lubrication data recording system was established, providing a reliable basis for equipment maintenance decisions and supporting full equipment lifecycle management and quality traceability.

[0075] The cloud-based collaborative communication interaction unit, based on the 5G-MEC edge computing node, conducts policy sharing communication between device groups. This unit interacts with the deep reinforcement learning model construction unit, blockchain data storage management unit, etc. to complete cloud-based data transmission and reception.

[0076] Specifically, the cloud-based collaborative communication and interaction unit, based on 5G-MEC edge computing nodes, enables high-speed communication and collaboration between the pumping unit lubrication system and the cloud platform, supporting policy sharing and data exchange among a fleet of millions of devices. This unit utilizes a distributed architecture, with edge computing nodes deployed on-site in the oilfield responsible for local data preprocessing and analysis, and a central server deployed in the cloud responsible for global data storage, analysis, and decision-making. The edge computing nodes and cloud servers are connected via a 5G network, enabling real-time data transmission and synchronization.

[0077] In terms of data processing, edge computing nodes possess powerful computing capabilities, enabling real-time processing and analysis of locally collected multimodal data. For example, they can extract features and perform fault diagnosis on vibration signals, and provide real-time monitoring and early warning of oil data. For simple decision-making tasks, edge computing nodes can handle them directly, reducing communication latency with the cloud. For complex analysis and decision-making tasks, edge computing nodes upload pre-processed data to cloud servers, leveraging the cloud's powerful computing resources for in-depth analysis. The cloud servers are responsible for storing and managing global data, establishing equipment archives and historical databases, and supporting predictive maintenance and optimization decisions.

[0078] The implementation of the cloud-based collaborative communication interaction unit involves 5G network deployment, edge computing platform construction, and cloud application development. The 5G network adopts the SA independent networking mode, providing low-latency, high-bandwidth communication services to meet real-time requirements. The edge computing platform uses industrial-grade servers equipped with high-performance CPUs and GPUs, supporting containerized deployment and microservices architecture. Cloud-based application development adopts a microservices architecture and DevOps development model to improve the scalability and maintainability of the system. The significance of this unit lies in the remote monitoring and collaborative management of the oil pumping unit lubrication system. Through the cloud platform, managers can understand the lubrication status of multiple oil pumping units in real time, formulate unified lubrication strategies, optimize resource allocation and collaborative management, and improve the overall efficiency and reliability of oilfield production.

[0079] Preferably, in the deep reinforcement learning model construction unit, the constructed deep reinforcement learning model uses the following formula to predict the oil film thickness:

[0080]

[0081] Among them, H t+1 represents the predicted oil film thickness at time t+1; is the eigenvector of the vibration spectrum of the pumping unit at time t after fast Fourier transform; M t is the concentration vector of metal particles in the oil at time t; T t is the temperature value at time t; P t is the pressure value at time t; α t , β t , γ t , δ t is a time-varying weight coefficient determined by the model through learning; θ is the parameter set of the model; f is a nonlinear mapping function based on deep learning, which realizes the prediction of oil film thickness by fusion processing of multimodal data.

[0082] Specifically, the oil film thickness prediction mechanism within the deep reinforcement learning model building unit dynamically predicts oil film thickness by integrating multiple parameters, including vibration spectrum characteristics, oil metal particle concentration, temperature, and pressure. After fast Fourier transform, the vibration spectrum characteristics reflect the operating status of various pumping unit components, such as gear mesh frequency and bearing failure frequency. The oil metal particle concentration directly reflects the degree of component wear. Temperature and pressure parameters are closely related to oil film formation and stability. By learning the complex nonlinear relationships between these parameters and oil film thickness, the model dynamically adjusts the weighting coefficients to meet the prediction requirements under different operating conditions. Compared to traditional single-parameter prediction, this multi-parameter fusion prediction method provides a more comprehensive and accurate reflection of oil film status, providing a key basis for the formulation of subsequent oil supply strategies. In practical applications, when the system predicts that the oil film thickness is approaching a critical value, it can adjust the oil supply in advance to avoid increased equipment wear due to oil film rupture.

[0083] Preferably, in the deep reinforcement learning model construction unit, the following reward function formula is adopted when optimizing lubrication parameters:

[0084]

[0085] Among them, R t represents the reward value at time t; E t is the energy consumption value of the pumping unit at time t; W t is the wear rate of key components of the pumping unit at time t; L t is the oil injection amount at time t; ω1, ω2, and ω3 are weight coefficients, and satisfy ω1+ω2+ω3=1. By adjusting these weight coefficients, an optimal balance among the three objectives of energy consumption, wear rate, and oil injection amount can be achieved.

[0086] Specifically, the reward function optimizes lubrication parameters by comprehensively considering three key indicators: energy consumption, wear rate, and oil injection volume. Energy consumption reflects the energy consumed during the operation of the pumping unit; reducing energy consumption can improve energy utilization efficiency; wear rate is directly related to the equipment's service life and maintenance costs; and oil injection volume affects the lubrication effect and lubricant consumption costs. A dynamic adjustment mechanism for the weight coefficients enables the model to balance these three objectives based on different application scenarios and priorities. For example, in energy-sensitive scenarios, the weight of the energy consumption indicator can be increased, making the model more inclined to choose low-energy lubrication solutions; in cases where equipment maintenance costs are high, the weight of the wear rate indicator can be increased to prioritize equipment reliability. This multi-objective optimization reward function design enables the system to maximize overall benefits while meeting lubrication needs.

[0087] Preferably, in the adaptive dynamic oil supply algorithm execution unit, the pulse high-pressure grease injection pressure adjustment formula is:

[0088] P inj =P0+k1·ΔV+k2·ΔT+k3·ΔM

[0089] Among them, P inj is the actual output grease injection pressure; P0 is the initial set pressure; ΔV is the difference in vibration spectrum characteristics between the current moment and the previous moment; ΔT is the temperature difference between the current moment and the previous moment; ΔM is the change in metal particle concentration in the oil between the current moment and the previous moment; k1, k2,

[0090] k3 is the pressure regulation coefficient, which is determined through experiments and model learning according to different operating conditions and equipment parameters of the pumping unit.

[0091] Specifically, the pulsed high-pressure grease injection pressure regulation mechanism uses an initial set pressure as its basis and dynamically adjusts based on changes in vibration spectrum characteristics, temperature, and metal particle concentration in the oil. Changes in vibration spectrum characteristics reflect changes in the equipment's operating status, such as load fluctuations or component failures; temperature changes directly affect the viscosity and fluidity of the lubricant; and changes in metal particle concentration in the oil indicate changes in equipment wear. The pressure regulation coefficient is determined through experiments and model learning based on different operating conditions and equipment parameters, ensuring that the grease injection pressure precisely matches the actual needs of the equipment. For example, during equipment startup or when load increases, the system automatically increases the grease injection pressure to ensure that the lubricant effectively penetrates the friction pair surfaces. During stable operation, the grease injection pressure is appropriately reduced to reduce energy consumption and lubricant waste. This dynamic pressure regulation mechanism improves lubrication effectiveness and reliability while reducing operating costs.

[0092] Preferably, in the adaptive dynamic oil supply algorithm execution unit, the oil injection amount adjustment formula is based on fuzzy PID control, as shown below:

[0093]

[0094] Wherein, ΔL is the adjustment amount of oil injection; e is the deviation between the current oil injection amount and the target oil injection amount; K p is the proportionality coefficient, K i is the integral coefficient, K d are differential coefficients, which are dynamically adjusted according to the load, operating time and other parameters of the pumping unit through fuzzy logic rules to achieve an accuracy adjustment of the oil injection volume of ±0.1ml.

[0095] Specifically, the oil injection quantity adjustment mechanism dynamically adjusts the parameters of the PID controller by monitoring the deviation between the current oil injection quantity and the target oil injection quantity and its rate of change in real time. The introduction of fuzzy logic rules enables the system to adaptively adjust control parameters based on complex operating conditions such as the pumping unit's load and operating time, achieving high-precision oil injection quantity adjustment. In actual application, when the system detects a deviation between the oil injection quantity and the target value, the fuzzy PID controller automatically adjusts the proportional, integral, and differential coefficients based on the magnitude and trend of the deviation. For example, when the deviation is large, the proportional coefficient is increased to speed up the response; when the deviation is small, the integral coefficient is increased to eliminate static errors. This intelligent adjustment mechanism can control the oil injection quantity within an accuracy of ±0.1ml, ensuring the stability and consistency of the lubrication effect while avoiding equipment failure and resource waste caused by excessive or insufficient oil injection.

[0096] Preferably, the deep reinforcement learning model construction unit and the adaptive dynamic fuel supply algorithm execution unit work together through the following information interaction mechanism:

[0097] A t =π θ (S t )

[0098] Among them, A t is the refueling action strategy output by the deep reinforcement learning model at time t; S t is the state vector collected by the multimodal composite parameter acquisition unit at time t and input into the deep reinforcement learning model after being processed by the data transmission and preprocessing adaptation unit; θ It is a policy network based on the parameters θ of the deep reinforcement learning model. The oil supply action strategy is generated through the network, and then the adaptive dynamic oil supply algorithm execution unit converts the strategy into a specific grease injection operation.

[0099] Specifically, the collaborative mechanism between the deep reinforcement learning model and the adaptive dynamic lubrication algorithm achieves close coordination between the two core units through the transmission of state vectors and the generation of action policies. The state vector, which contains vibration, oil level, temperature, and pressure data acquired by the multimodal composite parameter acquisition unit, is preprocessed and then fed into the deep reinforcement learning model. Based on this state information, the model generates an optimal lubrication action policy, such as grease injection time, pressure, and amount, through a policy network. The adaptive dynamic lubrication algorithm execution unit receives these policy instructions and converts them into specific grease injection operations. This collaborative mechanism enables the system to dynamically adjust the lubrication policy based on real-time monitoring data, forming a closed-loop intelligent control system. For example, if the system detects abnormal equipment vibration, the deep reinforcement learning model will promptly adjust the lubrication policy, increasing the injection amount or raising the pressure to improve lubrication conditions, thereby effectively reducing the occurrence of equipment failures.

[0100] Preferably, the data processing relationship formula between the multimodal composite parameter acquisition unit and the deep reinforcement learning model construction unit is:

[0101]

[0102] in, is the state vector input to the deep reinforcement learning model; is the original data vector collected by the multimodal composite parameter acquisition unit; φ is the data processing parameter set; g is the data processing function, which converts the original data into a state vector form suitable for deep reinforcement learning model input by performing feature extraction, normalization and other operations on the original data.

[0103] Specifically, the data processing relationship between the multimodal composite parameter acquisition unit and the deep reinforcement learning model construction unit involves a data processing function that converts the raw acquired data into a state vector suitable for model input. The raw data contains a significant amount of noise and redundant information, and the data formats and dimensions vary between sensors. The data processing function first extracts features from the raw data, screening out characteristic parameters closely related to the lubrication state. It then performs normalization to bring data of varying dimensions into the same scale range, improving the model's learning efficiency and accuracy. For example, vibration data is extracted using its peak value, root mean square value, and spectral characteristics. Temperature and pressure data are normalized to a distribution within the [0, 1] interval. This data processing mechanism ensures that the deep reinforcement learning model receives high-quality input data, thereby improving the model's prediction accuracy and generalization capabilities, providing strong support for accurately determining lubrication states and formulating optimization strategies.

[0104] Preferably, the adaptive dynamic oil supply algorithm execution unit, when considering the influence of the pumping unit's operating speed, uses the oil injection amount correction formula as follows:

[0105] L corr =L base (1+k v v)

[0106] Among them, L corr is the corrected oil filling amount; L base is the basic oil injection volume; v is the operating speed of the pumping unit; k v is the speed influence coefficient, which is determined through experiments and model training according to the mechanical structure and lubrication requirements of the pumping unit.

[0107] Specifically, an oil injection quantity correction mechanism was proposed based on the impact of the pumping unit's operating speed on the oil injection quantity. Operating speed is a key factor influencing lubrication effectiveness. As the pumping unit's operating speed changes, the lubricant distribution and oil film formation between the friction pairs also change. This correction mechanism uses a baseline oil injection quantity as a benchmark and dynamically adjusts the oil injection quantity based on the operating speed. The speed influence coefficient, determined through experiments and model training, reflects the extent to which operating speed affects lubrication requirements. In practical applications, as the pumping unit's operating speed increases, the lubricant between the friction pairs is easily thrown off, resulting in a decrease in oil film thickness. In this case, the system automatically increases the oil injection quantity to maintain optimal lubrication. Conversely, as the operating speed decreases, the oil injection quantity is appropriately reduced to avoid lubricant waste. This correction mechanism enables the system to better adapt to changes in pumping unit operating speed, improving the adaptability and accuracy of lubrication, and ensuring reliable operation of the equipment under various operating conditions.

[0108] Preferably, the deep reinforcement learning model construction unit adopts the following state transition probability formula when combining different working stages of the pumping unit:

[0109]

[0110] Among them, P(S t+1 ∣S t ,A t ) indicates that the state at time t is S t , perform action A t After that, it transfers to state S at time t+1 t+1 The probability of p i is the probability weight associated with different working stages of the pumping unit; f i (S t ,A t ) is the state transfer function corresponding to different working stages, which is obtained by learning the historical data of the pumping unit in different working stages to more accurately simulate the system state transfer process.

[0111] Specifically, a state transition probability model was proposed based on the characteristics of different operating stages of a pumping unit. The lubrication state changes and influencing factors vary across different operating stages, such as startup, stable operation, and shutdown. This model more accurately simulates the system state transition process by introducing probability weights and state transition functions associated with each operating stage. Each operating stage corresponds to a different state transition function, learned from historical data, that reflects the characteristics of the system state changes during that stage. The probability weights indicate the relative importance of each state transition function within the current operating stage. For example, during the startup phase, friction between equipment components is high, and lubrication state changes are more complex. Therefore, the state transition function corresponding to this phase has a larger weight. In contrast, during the stable operation phase, the lubrication state is relatively stable, and the corresponding state transition function has a larger weight. This staged state transition model improves the system's adaptability to different operating conditions and its prediction accuracy. It enables the deep reinforcement learning model to more accurately formulate lubrication strategies, effectively improving equipment lubrication management.

[0112] like Figure 2 As shown, the self-learning method of the lubrication state of the oil pumping unit based on multimodal sensing includes the following steps:

[0113] In the first step, a multi-modal composite parameter acquisition unit is used to collect parameters such as vibration spectrum, oil metal particles, temperature, and pressure during the operation of the pumping unit in real time.

[0114] In the second step, the collected multimodal parameter data is converted into a format, encoded, and adapted by the data transmission and preprocessing adaptation unit, and then transmitted to the deep reinforcement learning model construction unit;

[0115] In the third step, the deep reinforcement learning model construction unit uses the LSTM time series prediction module to learn the time series characteristics of the data based on the received data. Combined with the oil film thickness prediction algorithm, it constructs a deep reinforcement learning model to generate lubrication parameter optimization decisions.

[0116] In the fourth step, the adaptive dynamic oil supply algorithm execution unit receives the decision output by the deep reinforcement learning model construction unit and determines the specific oil injection pressure and oil injection volume adjustment strategy based on the pulsed high-pressure oil injection technology and fuzzy PID control logic;

[0117] In the fifth step, the decision result output and execution drive unit converts the optimization decision and oil supply strategy into control signals, driving the actuators of the oil pumping unit lubrication system to perform corresponding actions;

[0118] In the sixth step, the blockchain data evidence management unit records relevant information during the entire lubrication operation and securely stores and shares the data based on blockchain technology.

[0119] In the seventh step, the cloud-based collaborative communication interaction unit conducts strategy sharing and data interaction between the system and other pumping equipment based on the 5G-MEC edge computing node.

[0120] The intelligent lubrication decision-making system and method for oil pumps, based on multi-source data fusion, innovates the entire process from data collection, analysis and decision-making to execution and management, effectively addressing the shortcomings of traditional lubrication management. At the data collection level, traditional methods rely on single-point sensors or manual experience, making it difficult to fully capture lubrication status information. This system, however, uses a multimodal composite parameter acquisition unit to simultaneously collect multi-dimensional data such as vibration spectrum, oil metal particle concentration, temperature, and pressure. The high-precision vibration spectrum analysis module can monitor subtle vibration changes in component operation, the nano-level oil metal particle detection unit can capture early signs of wear, and the temperature-pressure dual-mode sensing node monitors the thermal pressure status of lubrication points in real time, obtaining comprehensive lubrication-related information to avoid potential faults caused by single and delayed data.

[0121] In the analysis and decision-making phase, traditional lubrication management uses fixed rules or simple models, which are unable to adapt to the complex and changing operating conditions of oil pumping units. This system leverages a deep reinforcement learning model, combined with LSTM time series prediction and oil film thickness prediction algorithms, to deeply mine multi-source data and accurately predict lubrication status trends. The adaptive dynamic oil supply algorithm makes decisions based on the model, combining pulsed high-pressure grease injection technology with fuzzy PID control logic to dynamically adjust the injection pressure and oil volume according to load, operating speed, and other operating conditions. For example, the system automatically increases the oil supply when the oil pumping unit starts under heavy load and reduces it during stable operation, eliminating the resource waste or lubrication insufficiency caused by traditional timed and quantitative oil supply, and achieving precise matching of lubrication strategies.

[0122] From a system management perspective, traditional data management is susceptible to tampering, difficult to share, and lacks coordination between devices. This system uses a blockchain data evidence management unit to encrypt, store, and share lubrication operation logs, ensuring data authenticity, reliability, and traceability. A cloud-based collaborative communication and interaction unit, based on 5G-MEC edge computing nodes, enables policy sharing and data interaction among millions of devices, supporting remote monitoring and collaborative management. The system establishes a closed-loop system from data collection and intelligent analysis to precise execution, effectively reducing equipment failure rates and maintenance costs, significantly improving the intelligence and refinement of pumping unit lubrication management, and meeting the needs of efficient and stable oilfield production.

[0123] In the description of the present invention, it should be noted that, unless otherwise expressly specified or limited, the terms "disposed," "installed," "connected," "connected," and "fixed" should be understood in a broad sense. For example, they may refer to fixed connections, detachable connections, or integral connections; they may refer to mechanical connections or electrical connections; they may refer to direct connections or indirect connections through an intermediate medium; and they may refer to internal communication between two components. Those skilled in the art will understand the specific meanings of the above terms in the present invention based on specific circumstances.

[0124] Although embodiments of the present invention have been shown and described, it will be understood by those skilled in the art that various equivalent changes, modifications, substitutions and variations may be made to these embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. The self-learning system of oil pumping unit lubrication status based on multi-modal sensing is characterized by: include: Multimodal composite parameter acquisition unit, data transmission and preprocessing adaptation unit, deep reinforcement learning model construction unit, adaptive dynamic fuel supply algorithm execution unit, decision result output and execution drive unit, blockchain data storage management unit, cloud collaborative communication interaction unit; The state composite parameter acquisition unit includes a high-precision vibration spectrum analysis module, a nano-level oil and metal particle detection module, and a temperature-pressure dual-mode sensing node. The unit adopts a double-layer sealed cavity design; The data transmission and preprocessing adaptation unit is used to encode and format-convert the data collected by the multimodal composite parameter acquisition unit, and transmit the data to subsequent units through a preset communication protocol; The deep reinforcement learning model construction unit integrates an LSTM time series prediction module to learn the time series characteristics of the collected data, and at the same time combines it with the oil film thickness prediction algorithm to build a model architecture based on deep reinforcement learning; The adaptive dynamic oil supply algorithm execution unit is equipped with pulsed high-pressure grease injection technology and combines fuzzy PID control logic to adjust the oil injection amount. The unit interacts with the deep reinforcement learning model construction unit and receives the oil supply strategy instructions output by the model.

2. The multi-modal sensing-based self-learning system for lubrication status of an oil pumping unit according to claim 1, characterized in that: The decision result output and execution drive unit converts the lubrication parameter optimization decision generated by the deep reinforcement learning model building unit and the oil supply strategy determined by the adaptive dynamic oil supply algorithm execution unit into a control signal that drives the action of the actuator of the oil pumping unit lubrication system; The blockchain data evidence management unit is used to record the lubrication operation log throughout the life cycle of the pumping unit and securely store and share data based on blockchain technology. This unit interacts with other units in the system to obtain relevant lubrication operation information. The cloud-based collaborative communication interaction unit is based on the 5G-MEC edge computing node to carry out policy sharing communication between the device groups. This unit interacts with the deep reinforcement learning model construction unit and the blockchain data evidence management unit to complete the cloud-based transmission and reception of data. The deep reinforcement learning model construction unit constructs a deep reinforcement learning model that uses the following formula to predict the oil film thickness: Among them, H t+1 represents the predicted oil film thickness at time t+1; is the eigenvector of the vibration spectrum of the pumping unit at time t after fast Fourier transform; M t is the concentration vector of metal particles in the oil at time t; T t is the temperature value at time t; P t is the pressure value at time t; α t , β t , γ t , δ t is a time-varying weight coefficient determined by the model through learning; θ is the parameter set of the model; f is a nonlinear mapping function based on deep learning, which realizes the prediction of oil film thickness by fusion processing of multimodal data.

3. The multi-modal sensing-based self-learning system for lubrication status of an oil pumping unit according to claim 1, characterized in that: The deep reinforcement learning model construction unit adopts the following reward function formula when optimizing lubrication parameters: Among them, R t represents the reward value at time t; E t is the energy consumption value of the pumping unit at time t; W t is the wear rate of key components of the pumping unit at time t; L t is the oil injection amount at time t; ω1, ω2, and ω3 are weight coefficients, and satisfy ω1+ω2+ω3=1. The weight coefficients are adjusted to achieve an optimal balance among the three objectives of energy consumption, wear rate, and oil injection amount.

4. The multi-modal sensing-based self-learning system for lubrication status of an oil pumping unit according to claim 1, characterized in that: The adaptive dynamic oil supply algorithm execution unit, the pulse high-pressure grease injection pressure adjustment formula is: P inj =P0+k1·Δv+k2·ΔT+k3·ΔM Among them, P inj is the actual output grease injection pressure; P0 is the initial set pressure; ΔV is the difference in vibration spectrum characteristics between the current moment and the previous moment; ΔT is the temperature difference between the current moment and the previous moment; ΔM is the change in metal particle concentration in the oil between the current moment and the previous moment; k1, k2, and k3 are pressure adjustment coefficients, which are determined through experiments and model learning according to different operating conditions and equipment parameters of the pumping unit.

5. The multi-modal sensing-based self-learning system for lubrication status of an oil pumping unit according to claim 1, characterized in that: The adaptive dynamic oil supply algorithm execution unit, the oil injection amount adjustment formula is based on fuzzy PID control, as shown below: Wherein, ΔL is the adjustment amount of oil injection; e is the deviation between the current oil injection amount and the target oil injection amount; K p is the proportionality coefficient, K i is the integral coefficient, K d are differential coefficients, which are dynamically adjusted according to the load and operating time parameters of the pumping unit through fuzzy logic rules.

6. The multi-modal sensing-based self-learning system for lubrication status of an oil pumping unit according to claim 1, characterized in that: The deep reinforcement learning model building unit and the adaptive dynamic fuel supply algorithm execution unit work together through the following information interaction mechanism: A t =π θ (S t ) Among them, A t is the refueling action strategy output by the deep reinforcement learning model at time t; S t is the state vector collected by the multimodal composite parameter acquisition unit at time t and input into the deep reinforcement learning model after being processed by the data transmission and preprocessing adaptation unit; θ It is a policy network based on the parameters θ of the deep reinforcement learning model. The oil supply action strategy is generated through the network, and then the adaptive dynamic oil supply algorithm execution unit converts the strategy into a specific grease injection operation.

7. The multi-modal sensing-based self-learning system for lubrication status of an oil pumping unit according to claim 1, characterized in that: The data processing relationship formula between the multimodal composite parameter acquisition unit and the deep reinforcement learning model construction unit is: in, is the state vector input to the deep reinforcement learning model; is the original data vector collected by the multimodal composite parameter acquisition unit; φ is the data processing parameter set; g is the data processing function, which converts the original data into a state vector form suitable for deep reinforcement learning model input by performing feature extraction and normalization operations on the original data.

8. The multi-modal sensing-based self-learning system for lubrication status of an oil pumping unit according to claim 1, characterized in that: When the adaptive dynamic oil supply algorithm execution unit takes into account the influence of the pumping unit's operating speed, the oil injection amount correction formula is: L corr =L base ·(1+k v ·v) Among them, L corr is the corrected oil filling amount; L base is the basic oil injection volume; v is the operating speed of the pumping unit; k v is the speed influence coefficient, which is determined through experiments and model training according to the mechanical structure and lubrication requirements of the pumping unit.

9. The multi-modal sensing-based self-learning system for lubrication status of an oil pumping unit according to claim 1, characterized in that: The deep reinforcement learning model construction unit adopts the following state transition probability formula when combining different working stages of the pumping unit: Among them, P(S t+1 ∣S t ,A t ) indicates that the state at time t is S t , perform action A t After that, it transfers to state S at time t+1 t+1 The probability of p i is the probability weight associated with different working stages of the pumping unit; f i (S t ,A t ) is the state transfer function corresponding to different working stages, which is obtained by learning the historical data of the pumping unit in different working stages to simulate the system state transfer process.

10. A self-learning method for lubrication status of an oil pumping unit based on multimodal sensing, characterized in that: The following steps are involved: In the first step, a multi-modal composite parameter acquisition unit is used to collect the vibration spectrum, oil metal particles, temperature and pressure parameters during the operation of the pumping unit in real time; In the second step, the collected multimodal parameter data is format converted and encoded by the data transmission and preprocessing adaptation unit, and then transmitted to the deep reinforcement learning model construction unit; In the third step, the deep reinforcement learning model construction unit uses the LSTM time series prediction module to learn the time series characteristics of the data based on the received data. Combined with the oil film thickness prediction algorithm, it constructs a deep reinforcement learning model to generate lubrication parameter optimization decisions. In the fourth step, the adaptive dynamic oil supply algorithm execution unit receives the decision output by the deep reinforcement learning model construction unit and determines the specific oil injection pressure and oil injection volume adjustment strategy based on the pulsed high-pressure oil injection technology and fuzzy PID control logic; In the fifth step, the decision result output and execution drive unit converts the optimization decision and oil supply strategy into control signals, driving the actuators of the oil pumping unit lubrication system to perform corresponding actions; In the sixth step, the blockchain data evidence management unit records relevant information during the entire lubrication operation and securely stores and shares the data based on blockchain technology. In the seventh step, the cloud-based collaborative communication interaction unit conducts strategy sharing and data interaction between the system and other pumping equipment based on the 5G-MEC edge computing node.

Images