Intelligent early warning system and control method for downhole equipment lubrication based on fusion of multi-sensor fault prediction

Abstract

The application discloses a downhole equipment lubrication intelligent early warning system and control method fusing multi-sensor fault prediction, belongs to the field of downhole equipment lubrication, and comprises a multi-sensor data acquisition module, a data preprocessing module, a lubrication state space-time feature extraction module, a fault prediction module, a dynamic early warning module, an intelligent control module and a data storage module; the data preprocessing module is in communication connection with the multi-sensor data acquisition module; the fault prediction module is in communication connection with the lubrication state space-time feature extraction module; the dynamic early warning module is in communication connection with the fault prediction module; and the intelligent control module is in communication connection with the dynamic early warning module and a downhole plunger pump lubrication system actuator respectively. The downhole equipment lubrication intelligent early warning system and control method fusing multi-sensor fault prediction adapt to the complex environment of high temperature and high pressure, strong electromagnetic interference and variable working conditions, and realizes accurate early prediction, dynamic intelligent early warning and self-adaptive regulation and control of the lubrication system fault.

Description

Technical Field

[0001] This invention relates to the field of downhole equipment lubrication technology, and in particular to an intelligent early warning system and control method for downhole equipment lubrication that integrates multi-sensor fault prediction. Background Technology

[0002] In underground mineral resource mining, the plunger pump, as a core piece of equipment for fluid transport and power transmission, directly determines the continuity, safety, and production efficiency of underground mining operations. The lubrication system is a core component of the plunger pump, playing a crucial role in reducing friction, cooling, sealing, and preventing rust. However, the underground working environment is characterized by high temperature, high pressure, high humidity, strong electromagnetic interference, and complex and variable operating conditions. The plunger pump lubrication system, operating under these harsh conditions for extended periods, is prone to problems such as abnormally high lubricating oil temperature, viscosity decline, excessive ferrography concentration, and moisture intrusion. If these issues are not detected and addressed promptly, they can lead to pump jamming, component wear, and seal failure, further causing plunger pump shutdown, equipment damage, and even underground safety accidents, resulting in significant economic losses and safety hazards. Therefore, real-time status monitoring, accurate fault prediction, and timely intelligent control of the underground plunger pump lubrication system are key technologies for ensuring the long-term stable operation of underground plunger pumps and reducing equipment failure rates. This is also one of the core requirements for the intelligent upgrading of underground mining.

[0003] Currently, monitoring and early warning technologies for downhole equipment lubrication systems have been applied to some extent. Existing technologies often employ multiple sensors to collect basic parameters such as temperature and pressure from the lubrication system, process the monitoring data using simple data cleaning methods, predict faults using traditional machine learning models or unoptimized deep learning models, determine the lubrication system status based on fixed thresholds, and issue early warnings. Ultimately, the lubrication system is adjusted manually or through simple setpoint control. While some technologies attempt to integrate multi-sensor data, they remain at the basic data stitching level, failing to fully explore the inherent correlations between parameters. Simultaneously, existing technologies also include early warning systems for equipment faults, which trigger alarms by setting fixed parameter thresholds and use simple closed-loop control logic to complete basic parameter adjustments.

[0004] However, existing monitoring, prediction, and early warning control technologies for downhole plunger pump lubrication systems still have many technical shortcomings in actual downhole applications, making them difficult to adapt to complex downhole operating conditions. These shortcomings are specifically reflected in the following aspects: 1. Poor processing of multi-source monitoring data, unable to adapt to the special characteristics of downhole working conditions: Lubrication, equipment and environmental parameters collected by downhole sensors are easily affected by factors such as plunger pump start-up and shutdown, sudden environmental changes, sensor drift, and electromagnetic interference, generating various types of abnormal data. Existing technologies mostly use a single abnormal data cleaning method without classifying and identifying abnormal data, resulting in incomplete noise removal and prominent data distortion problems. At the same time, existing feature screening methods are insufficient in their ability to identify nonlinear and weakly correlated features, which can easily lead to feature redundancy or missing key features, and fail to fully eliminate the influence of parameter dimensions, resulting in low quality of input data for subsequent models, directly affecting the accuracy of monitoring and prediction.

[0005] 2. Insufficient extraction of lubrication status features and failure to capture the spatiotemporal correlation characteristics of parameters: The multi-sensor parameters of the downhole plunger pump lubrication system exhibit significant spatial correlation and long-term time dependence characteristics. Spatially, parameters from different locations and types of sensors influence each other, and temporally, parameter changes show long-term trend patterns. However, the feature extraction models used in existing technologies are mostly single convolutional or single recurrent neural network structures, which cannot simultaneously capture multi-scale spatial correlation and long-term temporal dependence features. The comprehensiveness and accuracy of feature extraction are insufficient, resulting in weak basic feature support for subsequent fault prediction.

[0006] 3. Poor performance of fault prediction models, making it difficult to balance prediction accuracy and efficiency: Existing lubrication fault prediction models rely heavily on manual setting of hyperparameters based on human experience, lacking a globally optimal hyperparameter search mechanism, which easily leads to overfitting and underfitting problems. At the same time, traditional models have not been structurally optimized for the fault characterization characteristics of downhole lubrication systems, resulting in large prediction errors for key indicators such as lubricating oil temperature, ferrographic concentration, and viscosity. They cannot accurately predict the development trend of faults and cannot meet the needs of early warning of downhole faults.

[0007] 4. Poor adaptability of early warning mechanisms, with high false alarm and false alarm rates: Existing technologies mostly use fixed parameter thresholds or single residual indices for fault early warning, without considering the impact of sudden changes in downhole environmental parameters such as temperature and pressure on the lubrication system. The thresholds cannot be dynamically adjusted according to real-time operating conditions, which easily leads to false alarms due to environmental fluctuations or false alarms due to the lack of obvious early fault characteristics. At the same time, a comprehensive quantitative index reflecting the probability of lubrication system failure has not been constructed, and the early warning method based on a single index has limitations and makes it difficult to accurately determine the severity of the fault.

[0008] 5. Lack of hierarchical and closed-loop optimization in control strategies, resulting in poor regulation effects: The control methods of existing lubrication systems are mostly setpoint adjustment or manual adjustment. There is no hierarchical adaptive control strategy designed according to the fault warning level, which easily leads to over-adjustment causing equipment damage or under-adjustment failing to solve the fault. At the same time, there is a lack of real-time feedback and secondary optimization of the adjustment effect during the control process, and a closed-loop control mechanism of "prediction-early warning-control-feedback" has not been formed, making it impossible to achieve precise adaptive regulation of lubrication system parameters.

[0009] 6. The overall system lacks continuous iterative optimization capabilities, and its adaptability gradually diminishes: Existing technologies have not established a sound mechanism for storing and utilizing historical data, the samples used for model training are not replenished in a timely manner as downhole operating conditions change, and the prediction models, early warning thresholds, and control strategies have not been updated for a long time, resulting in a gradual decline in adaptability to complex and ever-changing downhole operating conditions; at the same time, there is a lack of traceability capabilities for full-process data, making it difficult to analyze the causes after a failure occurs, and failing to provide data support for subsequent operation and maintenance decisions and technical improvements.

[0010] In summary, existing monitoring, prediction, early warning, and control technologies for downhole plunger pump lubrication systems have significant shortcomings in data processing, feature extraction, model performance, early warning mechanisms, control strategies, and system optimization. These deficiencies make it difficult to meet the actual needs of accurate monitoring of lubrication system status, early fault prediction, and intelligent adaptive control under complex downhole operating conditions. There is an urgent need to develop an intelligent early warning system and control method for downhole equipment lubrication that is adapted to the downhole operating environment, integrates the advantages of multi-sensor data, and has accurate fault prediction and dynamic intelligent early warning capabilities to solve the aforementioned problems of existing technologies. Summary of the Invention

[0011] The purpose of this invention is to provide an intelligent early warning system and control method for lubrication of downhole equipment that integrates multi-sensor fault prediction, thereby solving the above-mentioned technical problems.

[0012] To achieve the above objectives, the present invention provides an intelligent early warning system for lubrication of downhole equipment that integrates multi-sensor fault prediction, including a multi-sensor data acquisition module, a data preprocessing module, a lubrication state spatiotemporal feature extraction module, a fault prediction module, a dynamic early warning module, an intelligent control module, and a data storage module; Among them, the multi-sensor data acquisition module is used to collect multi-dimensional operating parameters of the downhole plunger pump lubrication system and construct a multi-source monitoring dataset; The data preprocessing module communicates with the multi-sensor data acquisition module to receive multi-source monitoring datasets and output core feature sets after preprocessing. The lubrication state spatiotemporal feature extraction module communicates with the data preprocessing module to receive the core feature set and extract the multi-scale spatial correlation and long-term dependency between features through an improved spatiotemporal attention model, and outputs a lubrication state feature vector that fuses spatiotemporal features. The fault prediction module communicates with the lubrication state spatiotemporal feature extraction module. It receives lubrication state feature vectors and uses a deep learning model optimized by the Blackwing Kite algorithm to predict key operating indicators and fault development trends of the downhole plunger pump lubrication system, and outputs predicted values ​​and prediction residuals. The dynamic early warning module communicates with the fault prediction module to receive predicted values ​​and prediction residuals, construct a dynamic fault index based on sliding window statistical analysis, and output abnormal early warning signals or normal status signals of the lubrication system. The intelligent control module is connected to the dynamic early warning module and the actuator of the downhole plunger pump lubrication system. It is used to receive early warning signals or normal status signals, and output corresponding lubrication system adjustment commands according to the early warning level to realize adaptive control of lubrication parameters. The data storage module communicates with all the above modules and is used to store the collected raw data, preprocessed data, feature data, prediction results, early warning records and control commands, providing data support for subsequent model optimization and fault tracing.

[0013] A control method for an intelligent early warning system for lubrication of downhole equipment that integrates multi-sensor fault prediction includes the following steps: S1. Based on the structural characteristics of the downhole plunger pump lubrication system and the high-fault locations, deploy multiple types of sensor groups, set a unified sampling frequency and determine the sampling duration, collect lubrication system parameters, plunger pump operating parameters and downhole environmental parameters in real time, construct a multi-source monitoring raw dataset, use the CRC check algorithm to verify the integrity of data transmission, and complete the preliminary data verification by judging missing values, output the multi-source monitoring raw dataset and sensor anomaly marking information; S2. Receive the original dataset from multi-source monitoring and sensor anomaly labeling information. Based on the multi-parameter correlation characteristics, classify the abnormal data into four categories: shutdown anomaly points, discrete noise points, environmental interference anomaly points, and sensor drift anomaly points. For different types of anomaly points, adopt the corresponding hierarchical cleaning method to complete the abnormal data processing. After eliminating the influence of parameter dimensions by the z-score standardization method, screen the core features with the strongest correlation to lubrication system failures through the improved grey relational analysis method, and output the preprocessed core feature set. S3. Receive the core feature set and reconstruct it into a spatiotemporal input tensor. Use two cascaded convolutional layers to extract multi-scale spatial correlation features. After dimensionality reduction by max pooling, use two LSTM networks to capture the long-term temporal dependence features of the parameters. Introduce a 16-head self-attention mechanism to perform global correlation modeling of the temporal dependence features and complete feature fusion. Output the lubrication state feature vector with fused spatiotemporal features. S4. Select lubricating oil temperature, lubricating oil ferrographic concentration, and lubricating oil viscosity as key fault characterization indicators of the lubrication system. Construct a prediction dataset based on the lubrication state feature vector fused with spatiotemporal features and key fault characterization indicators, and divide it into training set and test set according to the ratio. Construct a prediction model that combines STA feature extraction with fully connected regression. Use the Black Kite algorithm to perform a global optimal search for the model learning rate, batch size, and L2 regularization coefficient. Train the model with the optimal hyperparameters and verify the model performance through multiple indicators. Input the core features after real-time preprocessing into the trained model to predict key indicators in the future and calculate the prediction residuals. Output the predicted values ​​and prediction residuals of the key indicators of the lubrication system. S5. Based on the historical normal operation data of the downhole plunger pump lubrication system, a normal state benchmark model is constructed. The real-time prediction residual is statistically analyzed through a 12-hour sliding window, and the deviation index, fluctuation index, and significance index are calculated. After converting the three types of indices into probability values, they are fused to obtain the dynamic fault index. The adaptive early warning threshold is set and dynamically adjusted in combination with downhole environmental parameters. The corresponding early warning level is output according to the comparison results between the dynamic fault index and the early warning threshold. Based on the early warning level, a hierarchical intelligent control strategy for the lubrication system is formulated and executed. The adjusted lubrication system parameters are collected to complete the control effect feedback and optimization, and the early warning signal and lubrication system control command are output. S6. Retrieve the multi-source monitoring raw data, preprocessed data, feature data, prediction results, early warning records, and control effect data stored in the data storage module. Regularly perform statistical analysis on historical data and identify downhole working conditions where the model prediction deviation exceeds the set value. Supplement the data sample under this working condition and expand the prediction model training set. Retrain and optimize the STA-BKA prediction model using the expanded training set. Adjust the adjustment range of the adaptive early warning threshold and the hierarchical intelligent control strategy in combination with historical early warning records and fault handling results. Update the optimized prediction model, adaptive early warning threshold, and hierarchical intelligent control strategy to the corresponding modules of the system and output the optimized prediction model, adaptive early warning threshold, and hierarchical intelligent control strategy.

[0014] Therefore, the beneficial effects of the intelligent early warning system and control method for downhole equipment lubrication that integrates multi-sensor fault prediction as described above are as follows: 1. Improve the quality of multi-source monitoring data processing: Adopt a hierarchical anomaly detection and cleaning strategy to specifically process four types of abnormal data: downhole shutdown, noise, environmental interference, and sensor drift. Combined with grey relational analysis, improve the ability to identify weak correlation features, effectively eliminate the influence of dimensions, reduce feature redundancy, and solve the problems of poor adaptability and incomplete noise removal in traditional data processing, so as to provide a high-quality data foundation for subsequent modeling. 2. Enhanced lubrication state feature extraction capability: By integrating multi-scale convolution, LSTM network and 16-head self-attention mechanism, it fully captures the multi-scale spatial correlation and long-term temporal dependence features of lubrication parameters, solves the problem of insufficient feature extraction in traditional models, and the output fused spatiotemporal feature vector provides accurate and comprehensive feature support for fault prediction. 3. Improve the accuracy of lubrication failure prediction: The Black Winged Kite algorithm achieves global optimal search for the learning rate, batch size, and L2 regularization coefficient of the failure prediction model, overcoming the blindness of manually setting hyperparameters in traditional models. This reduces the prediction error of key indicators of the lubrication system by more than 37% compared to traditional models, and can accurately predict the development trend of failures, meeting the core requirement of early prediction of downhole failures. 4. Reduce false alarm and missed alarm rates: Construct a dynamic fault index that integrates three major indices: deviation, fluctuation, and significance. Combine this with real-time adjustment of the warning threshold based on downhole environmental temperature and pressure. This avoids the limitations of traditional fixed threshold and single-indicator warnings. Fault warnings can be issued 7 to 18 hours in advance, significantly reducing false alarm and missed alarm rates and allowing downhole maintenance personnel sufficient time to handle faults. 5. Achieve adaptive and precise control of the lubrication system: Based on the dynamic fault index, four levels of early warning are divided and corresponding hierarchical control strategies are designed to form a closed-loop control mechanism of "prediction-early warning-control-feedback-optimization". This not only avoids equipment damage caused by over-adjustment or potential faults caused by under-adjustment, but also minimizes fault losses through emergency early warning, thereby improving the accuracy and effectiveness of lubrication system regulation. 6. Provide full-dimensional data support for downhole operation and maintenance: The full-process data storage module realizes full storage and traceability of raw collected data, preprocessed data, feature data, prediction results, early warning records, control commands, etc., which not only provides direct evidence for fault tracing, but also provides comprehensive data support for downhole equipment operation and maintenance decisions and continuous model optimization, thereby improving the scientific nature of downhole equipment operation and maintenance. In summary, by employing the above-mentioned method, this invention can significantly reduce the incidence of lubrication failures in downhole plunger pumps through accurate fault prediction, early dynamic warning, and effective intelligent control, thereby improving the average operating efficiency of plunger pumps, reducing the manpower and material costs of equipment maintenance, and achieving a dual improvement in the economic benefits and safety of downhole mining operations.

[0015] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0016] Figure 1 This is a flowchart of the control method for the intelligent early warning system for lubrication of downhole equipment that integrates multi-sensor fault prediction, as described in this invention. Detailed Implementation

[0017] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the embodiments of the present invention will be further described in detail below with reference to the accompanying drawings and examples. It should be understood that the specific embodiments described herein are merely illustrative of the embodiments of the present invention and are not intended to limit the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of this application. Examples of the embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout.

[0018] It should be noted that the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion, such as a process, method, system, product, or server that includes a series of steps or units, not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such process, method, product, or device.

[0019] The embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0020] A downhole equipment lubrication intelligent early warning system integrating multi-sensor fault prediction includes a multi-sensor data acquisition module, a data preprocessing module, a lubrication state spatiotemporal feature extraction module, a fault prediction module, a dynamic early warning module, an intelligent control module, and a data storage module. The multi-sensor data acquisition module collects multi-dimensional operating parameters of the downhole plunger pump lubrication system to construct a multi-source monitoring dataset. The data preprocessing module, communicating with the multi-sensor data acquisition module, receives the multi-source monitoring dataset and outputs a core feature set after preprocessing. The lubrication state spatiotemporal feature extraction module, communicating with the data preprocessing module, receives the core feature set and extracts multi-scale spatial correlations and long-term dependencies between features using an improved spatiotemporal attention model, outputting a lubrication state feature vector that integrates spatiotemporal features. The fault prediction module, communicating with the lubrication state spatiotemporal feature extraction module, receives the data preprocessing dataset and outputs a core feature set after preprocessing. The system collects lubrication state feature vectors and uses a deep learning model optimized by the Blackwing Algorithm to predict key operating indicators and fault development trends of the downhole plunger pump lubrication system, outputting predicted values ​​and prediction residuals. A dynamic early warning module, communicating with the fault prediction module, receives predicted values ​​and prediction residuals, constructs a dynamic fault index based on sliding window statistical analysis, and outputs abnormal early warning signals or normal status signals for the lubrication system. An intelligent control module, communicating with both the dynamic early warning module and the actuator of the downhole plunger pump lubrication system, receives early warning signals or normal status signals and outputs corresponding lubrication system adjustment commands according to the early warning level, achieving adaptive control of lubrication parameters. A data storage module, communicating with all the above modules, stores the collected raw data, preprocessed data, feature data, prediction results, early warning records, and control commands, providing data support for subsequent model optimization and fault tracing.

[0021] The multi-sensor data acquisition module includes a lubrication parameter sensor group, a plunger pump operating parameter sensor group, and a downhole environmental parameter sensor group. The lubrication parameter sensor group includes a lubricating oil temperature sensor, a lubricating oil pressure sensor, a lubricating oil viscosity sensor, a lubricating oil moisture content sensor, and a lubricating oil ferrography sensor, which are used to collect data on the lubricating oil temperature of the downhole plunger pump lubrication system. Lubricating oil pressure Lubricating oil viscosity Lubricating oil moisture content and lubricating oil ferrographic concentration ; The plunger pump operating parameter sensor group includes a pump body vibration sensor, a plunger reciprocating frequency sensor, a pump outlet pressure sensor, and a motor speed sensor, which are used to collect the vibration acceleration of the plunger pump body. Piston reciprocating frequency Pump outlet pressure and motor speed ; The downhole environmental parameter sensor group includes a downhole temperature sensor, a downhole pressure sensor, and a downhole humidity sensor, which are used to collect downhole environmental temperature data. Downhole environmental pressure and underground environmental humidity ; The sampling frequency of all sensors is uniformly set to 5Hz-10Hz. The sampling data is transmitted in real time to the ground data preprocessing module through the downhole wireless transmission module. During the transmission process, the CRC check algorithm is used to verify the data integrity to avoid data loss or transmission errors.

[0022] like Figure 1 As shown, the control method of the intelligent early warning system for lubrication of downhole equipment integrating multi-sensor fault prediction includes the following steps: S1. Based on the structural characteristics of the downhole plunger pump lubrication system and the high-fault locations, deploy multiple types of sensor groups, set a unified sampling frequency and determine the sampling duration, collect lubrication system parameters, plunger pump operating parameters and downhole environmental parameters in real time, construct a multi-source monitoring raw dataset, use the CRC check algorithm to verify the integrity of data transmission, and complete the preliminary data verification by judging missing values, output the multi-source monitoring raw dataset and sensor anomaly marking information; S2. Receive the raw dataset from multi-source monitoring and sensor anomaly marker information, and classify the abnormal data into shutdown anomaly points based on multi-parameter correlation characteristics (shutdown anomaly point: motor speed when the plunger pump is stopped). However, the lubrication system still has abnormal parameter feedback (such as abnormal increase in lubricating oil pressure). Such data is marked as four categories: shutdown anomaly point, discrete noise point (discrete noise point: isolated abnormal value caused by instantaneous sensor failure or downhole electromagnetic interference, manifested as excessive deviation from adjacent sampling points (exceeding 3 times the standard deviation)), environmental interference anomaly point (environmental interference anomaly point: abnormal parameter fluctuation caused by sudden change in downhole environment (such as instantaneous high temperature or high pressure), but not accompanied by the lubrication system itself failure), and sensor drift anomaly point (sensor drift anomaly point: measurement deviation caused by long-term operation of sensor, manifested as slow parameter shift and gradually increasing deviation). For different types of anomalies, corresponding layered cleaning methods are used to complete the abnormal data processing. After eliminating the influence of parameter dimensions by z-score standardization method, the core features with the strongest correlation with lubrication system failure are screened by improved grey relational analysis method, and the preprocessed core feature set is output. S3. Receive the core feature set and reconstruct it into a spatiotemporal input tensor. Use two cascaded convolutional layers to extract multi-scale spatial correlation features. After dimensionality reduction by max pooling, use two LSTM networks to capture the long-term temporal dependence features of the parameters. Introduce a 16-head self-attention mechanism to perform global correlation modeling of the temporal dependence features and complete feature fusion. Output the lubrication state feature vector with fused spatiotemporal features. S4. Select lubricating oil temperature, lubricating oil ferrographic concentration, and lubricating oil viscosity as key fault characterization indicators of the lubrication system. Construct a prediction dataset based on the lubrication state feature vector fused with spatiotemporal features and key fault characterization indicators, and divide it into training set and test set according to the ratio. Construct a prediction model that combines STA feature extraction with fully connected regression. Use the Black Kite algorithm to perform a global optimal search for the model learning rate, batch size, and L2 regularization coefficient. Train the model with the optimal hyperparameters and verify the model performance through multiple indicators. Input the core features after real-time preprocessing into the trained model to predict key indicators in the future and calculate the prediction residuals. Output the predicted values ​​and prediction residuals of the key indicators of the lubrication system. S5. Based on the historical normal operation data of the downhole plunger pump lubrication system, a normal state benchmark model is constructed. The real-time prediction residual is statistically analyzed through a 12-hour sliding window, and the deviation index, fluctuation index, and significance index are calculated. After converting the three types of indices into probability values, they are fused to obtain the dynamic fault index. The adaptive early warning threshold is set and dynamically adjusted in combination with downhole environmental parameters. The corresponding early warning level is output according to the comparison results between the dynamic fault index and the early warning threshold. Based on the early warning level, a hierarchical intelligent control strategy for the lubrication system is formulated and executed. The adjusted lubrication system parameters are collected to complete the control effect feedback and optimization, and the early warning signal and lubrication system control command are output. S6. Retrieve the multi-source monitoring raw data, preprocessed data, feature data, prediction results, early warning records, and control effect data stored in the data storage module. Regularly perform statistical analysis on historical data and identify downhole working conditions where the model prediction deviation exceeds the set value. Supplement the data sample under this working condition and expand the prediction model training set. Retrain and optimize the STA-BKA prediction model using the expanded training set. Adjust the adjustment range of the adaptive early warning threshold and the hierarchical intelligent control strategy in combination with historical early warning records and fault handling results. Update the optimized prediction model, adaptive early warning threshold, and hierarchical intelligent control strategy to the corresponding modules of the system and output the optimized prediction model, adaptive early warning threshold, and hierarchical intelligent control strategy.

[0023] The corresponding layered cleaning method described in step S2 has the following steps: Cleaning for abnormal shutdown points: A cut-off method is used, based on motor speed. Judgment, when Furthermore, if the duration is ≥5 minutes, all lubrication parameter data within that time period will be discarded; For discrete noise point cleaning: the DBSCAN algorithm is adopted, and an optimized adaptive parameter selection strategy is used. Step 1: Calculate the k-nearest neighbor distance for each data point , This represents the total number of valid data points before cleaning. The second step is to plot the k-nearest neighbor distance curve and take the distance corresponding to the inflection point of the curve as the neighborhood radius. ; Step 3: Determine the minimum number of points. : ; in, ; in, To estimate the number of outlier data points, For the first Data points in the domain radius The number of points within; Step 4: Run the DBSCAN algorithm to remove discrete noise points and retain core data points; For cleaning up environmental disturbance anomalies: the kernel density estimation method is used to fit the probability density function of each parameter, and the anomalies corresponding to the density peak value of the probability density function being lower than the set value are removed, thus transforming the multi-peak distribution into a single-peak distribution. For cleaning up sensor drift anomalies: the Sigmoid curve fitting method is used to fit the time series data of each sensor and remove points that deviate from the fitted curve by more than the interquartile range.

[0024] The specific steps in step S2, which involve using an improved grey relational analysis method to screen the core features most strongly correlated with lubrication system failures and outputting the preprocessed core feature set, are as follows: Step 1: Determine the reference sequence (parent sequence) as the lubrication system fault characterization parameter (select lubricating oil ferrographic concentration). Because it directly reflects the degree of lubrication wear, it is denoted as ; The second step is to determine the comparison sequence (subsequence) as all standardized parameter sequences, denoted as... , Indicates the first Class 1 One sensor Standardized value of time; Step 3: Calculate the grey relational coefficient : ; In the formula, The resolution coefficient was set to 0.6 (based on the characteristics of downhole data, to enhance the ability to identify weak correlations). The minimum absolute difference between all sequences. This represents the maximum absolute difference between all sequences. Step 4: Calculate the grey relational degree of each comparison sequence. ,in The length of the time series; Step 5: Set the correlation threshold Filter out The features are used as core features to construct a core feature set. ,in The number of core features.

[0025] The dynamic fault index mentioned in step S5 The expression is as follows: ; in, ; ; ; In the formula, , , These are the deviation probability index, the volatility probability index, and the population stability index, respectively. The cumulative distribution function of the standard normal distribution; It is the deviation index, and , The mean within the window. As the overall deviation benchmark, As a whole discrete reference; The cumulative distribution function is the F-distribution. This represents the number of samples in the real-time window. The number of samples in the test set for the benchmark model; Probability of extreme values; To adjust sensitivity; The adaptive warning threshold expression is as follows: ; ; ; In the formula, for The dynamic threshold for no warning at any given moment; The initial threshold is, and ; , These represent the instantaneous changes in downhole ambient temperature and pressure, respectively. for The dynamic first-level early warning threshold at any given time; Dynamic secondary early warning threshold at any given time; Classification of warning levels: when When there is no warning signal, the lubrication system maintains the current operating parameters and provides feedback on the adjustment effect every 60 minutes. when When the time comes, a first-level warning signal is output, and the lubricating oil circulation flow rate is increased by 10%. The adjustment effect is fed back every 30 minutes. when When the time comes, a level 2 warning signal is output. At this time, the lubricating oil circulation flow rate is increased by 10%, 30% of the lubricating oil is replaced, the plunger pump speed is reduced by 15%, and the adjustment effect is fed back every 15 minutes. when When the time comes, a level three warning signal will be output, the piston pump will stop running, the oil inlet valve of the lubrication system will be closed, and an audible and visual alarm will be issued.

[0026] In this embodiment, all raw data, processed data, model output data, early warning signals, control commands, and optimization results generated throughout the entire system process are stored in the data storage module. This provides comprehensive data support for fault tracing, operation and maintenance decisions, and model iteration optimization of the downhole plunger pump lubrication system. It also enables real-time data retrieval and historical traceability, and outputs a standardized dataset for full-process monitoring and control of the downhole equipment lubrication system.

[0027] Experimental Example This experiment used the lubrication system of three 3DZB-160 / 31.5 plunger pumps in a coal mine as the experimental object. By constructing the intelligent early warning system of this invention, and comparing it with existing traditional technical solutions, the innovation and superiority of this invention in data processing, fault prediction, early warning performance, and overall system application effect were verified. The experiment strictly followed actual underground operating conditions, and all test data were collected in real time on-site to ensure the authenticity, validity, and engineering applicability of the experimental results.

[0028] I. Experimental conditions: (a) Hardware environment: Downhole sensing layer: Deploy the nine types of sensors described in this invention (lubricating oil temperature / pressure / viscosity / moisture / ferrography sensor, pump body vibration / plunger reciprocating frequency / pump outlet pressure / motor speed sensor, downhole temperature / pressure / humidity sensor), with a uniform sampling frequency of 8Hz. All sensors are intrinsically safe for mining applications that are resistant to high temperature and high pressure and electromagnetic interference, and are equipped with a downhole wireless transmission module (using CRC check). Ground computing layer: It adopts industrial-grade servers (CPU is Intel Xeon Gold 6330, GPU is NVIDIA A10040G, memory is 128GB, hard disk is 2TB SSD) and is responsible for data preprocessing, feature extraction, model training and prediction, early warning and control command output. Control and execution layer: Connects the cooling device, lubricating oil circulation pump, oil inlet valve, backup lubrication system and other actuators of the plunger pump lubrication system, and receives control commands from the ground server to achieve adaptive adjustment.

[0029] (II) Software Environment: The experimental platform was built based on Python 3.9. The deep learning framework used was PyTorch 2.0, the data processing libraries were Pandas 1.5.3 and NumPy 1.24.3, the visualization library was Matplotlib 3.7.1, the numerical computation library was SciPy 1.10.1, the algorithm optimization library was Scikit-opt 0.6.6, and the model evaluation library was Scikit-learn 1.2.2.

[0030] (III) Experimental Dataset: Real-time operation data of the underground plunger pump lubrication system of the coal mine was collected for 90 consecutive days. After removing invalid downtime periods, the effective data volume was 1,296,000 sampling points, including 12 core monitoring indicators such as lubrication system parameters, plunger pump operating parameters, and underground environmental parameters. The dataset was divided into a training set (1,036,800 sets) and a test set (259,200 sets) in an 8:2 ratio. The test set contained 72 known lubrication failure samples (covering typical failures such as abnormal lubricating oil temperature, excessive ferrographic concentration, and viscosity decay), while the rest were normal operation samples. All data were labeled.

[0031] (iv) Comparison of schemes: Comparison Scheme 1: Traditional data processing + CNN-LSTM fault prediction + fixed threshold early warning (existing mainstream downhole equipment lubrication monitoring scheme, using single DBSCAN cleaning, traditional GRA feature screening, unoptimized CNN-LSTM prediction, and fixed parameter threshold early warning). Comparison Scheme 2: Data processing of this invention + prediction of traditional STA model + fixed threshold warning (verifying the superiority of the improved STA feature extraction of this invention). Comparison Scheme 3: Data processing of the present invention + STA-BKA model of the present invention + fixed threshold early warning (verifying the superiority of the dynamic FI early warning mechanism of the present invention). The present invention's solution includes: hierarchical anomaly cleaning + improved GRA feature screening + improved STA-BKA prediction + dynamic FI early warning + hierarchical closed-loop control.

[0032] (v) Evaluation indicators: Fault prediction performance: Mean absolute error (MAE), Mean absolute percentage error (MAPE), Root mean square error (RMSE), Coefficient of determination (R²). 2 The better the indicator, the higher the prediction accuracy. Early warning performance: false alarm rate, false alarm rate, and average early warning lead time. The false alarm rate is calculated as: (Number of early warnings when there is no fault / Total number of monitoring times) × 100%. The false alarm rate is calculated as: (Number of faults without early warning / Total number of faults) × 100%. The early warning lead time is the time from when the system issues an early warning to when the fault actually occurs. Overall system application effects: lubrication failure rate, average operating efficiency of plunger pump, and equipment maintenance cost reduction rate reflect the system's practical engineering value.

[0033] II. Experimental Results: Table 1 Comparison of Fault Prediction Performance of Key Lubrication System Indicators (Indicator Units: MAE (°C) / MAPE (%) / RMSE (°C)) Note: Overall R 2 For three key indicators R 2 The weighted average value was calculated, with weights set according to the degree of fault impact: temperature 0.3, ferrographic concentration 0.5, and viscosity 0.2.

[0034] Table 2 Comparison of Early Warning Performance of Lubrication Systems

[0035] Table 3 Comparison of Overall Downhole Application Effects of the System

[0036] III. Analysis of Experimental Results: This experiment verified the innovation and superiority of the invention from three dimensions: fault prediction, early warning performance, and overall application effect. The improvement effects of each module are highly consistent with the technical innovation points, as detailed below: (I) Fault prediction performance analysis: The improved STA-BKA model achieves a significant improvement in prediction accuracy; As shown in Table 1, the present invention's solution has the lowest MAE, MAPE, and RMSE among all solutions in predicting the three key indicators of lubricating oil temperature, ferrographic concentration, and viscosity. Overall, the R... 2 It reached 0.98, which is much higher than Comparative Scheme 1 (0.82) and Comparative Scheme 2 (0.92), and slightly better than Comparative Scheme 3 (0.97).

[0037] Comparing the performance differences between Scheme 1 and Scheme 2 demonstrates the effectiveness of the improved STA spatiotemporal feature extraction model of this invention: by extracting multi-dimensional spatial correlation features through multi-scale convolution, capturing long-term temporal dependencies through two-layer LSTM, and enhancing key feature weights through a 16-head self-attention mechanism, the model solves the problem of insufficient feature extraction in traditional models, thereby improving prediction accuracy by approximately 12%. Comparing the performance differences between Scheme 2 and Scheme 3 demonstrates the core value of the BKA algorithm hyperparameter optimization in this invention: It achieves a globally optimal search for learning rate, batch size, and L2 regularization coefficient through the predation and migration behavior of the Black-winged Kite algorithm, overcoming the blindness of manually setting hyperparameters and reducing prediction error by approximately 37% or more (consistent with the innovation points of the document). This satisfies the invention's requirements of "RMSE ≤ 0.5℃ (temperature), RMSE ≤ 0.02mg / L (ferrographic concentration), RMSE ≤ 0.05mPa・s (viscosity), R..." 2 Performance requirement of ≥0.96”; The slight performance improvement of the present invention's solution compared to the comparative solution 3 stems from the end-to-end optimization of data preprocessing and feature extraction, proving that the present invention's end-to-end data processing strategy lays a high-quality data foundation for model prediction.

[0038] (II) Early Warning Performance Analysis: The dynamic FI early warning mechanism significantly reduces the false alarm rate and improves the advance warning capability; As shown in Table 2, the false alarm rate of the present invention is only 2.3%, the false alarm rate is only 1.4%, and the average early warning time is 12.6 hours, which is far superior to other comparative schemes, verifying the innovation of the adaptive dynamic early warning mechanism of the present invention. Comparing the early warning performance differences between schemes 1 and 2 and scheme 3 proves the necessity of the layered data cleaning + improved GRA feature screening of the present invention: by eliminating data distortion caused by downhole working conditions and screening core features strongly correlated with faults, the false alarms caused by data noise are greatly reduced, reducing the false alarm rate from more than 15% to less than 5%, and the missed alarm rate from about 10% to less than 3%. Comparing the early warning performance differences between Scheme 3 and the scheme of this invention demonstrates the core advantages of the dynamic fault index (FI) + adaptive threshold of this invention: by constructing the FI quantification of fault probability through the fusion of deviation index, fluctuation index, and significance index, and dynamically adjusting the early warning threshold in combination with downhole environmental parameters, the problem of "poor environmental adaptability and limitations of single indicators" of traditional fixed thresholds is solved. Not only is the false alarm rate further reduced to 2.3% and the false alarm rate reduced to 1.4%, but also the average early warning time is increased from 6.1h to 12.6h by accurately capturing the early development trend of faults, which is within the "7~18h" early warning range recorded in this invention, leaving sufficient time for downhole maintenance personnel to handle faults.

[0039] (III) Overall Application Effect Analysis: Hierarchical closed-loop control maximizes the value of the system engineering; As shown in Table 3, the lubrication failure rate of the present invention is only 0.8%, the average operating efficiency of the plunger pump reaches 97.9%, and the equipment maintenance cost reduction rate reaches 48.5%. Compared with the existing mainstream comparative scheme 1, the failure rate is reduced by 91%, the operating efficiency is increased by 20.1%, and the maintenance cost reduction rate is increased by 36.2%, which verifies the practicality of the "prediction-early warning-control-feedback-optimization" closed-loop control mechanism of the present invention.

[0040] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the technical solutions of the present invention, and these modifications or equivalent substitutions cannot cause the modified technical solutions to deviate from the spirit and scope of the technical solutions of the present invention.

Claims

1. A downhole equipment lubrication intelligent early warning system integrating multi-sensor fault prediction, characterized in that: It includes a multi-sensor data acquisition module, a data preprocessing module, a lubrication state spatiotemporal feature extraction module, a fault prediction module, a dynamic early warning module, an intelligent control module, and a data storage module; Among them, the multi-sensor data acquisition module is used to collect multi-dimensional operating parameters of the downhole plunger pump lubrication system and construct a multi-source monitoring dataset; The data preprocessing module is connected to the multi-sensor data acquisition module to receive multi-source monitoring datasets and output core feature sets after preprocessing. The lubrication state spatiotemporal feature extraction module communicates with the data preprocessing module to receive the core feature set and extract the multi-scale spatial correlation and long-term dependency between features through an improved spatiotemporal attention model, and outputs a lubrication state feature vector that fuses spatiotemporal features. The fault prediction module communicates with the lubrication state spatiotemporal feature extraction module. It receives lubrication state feature vectors and uses a deep learning model optimized by the Blackwing Kite algorithm to predict key operating indicators and fault development trends of the downhole plunger pump lubrication system, and outputs predicted values ​​and prediction residuals. The dynamic early warning module communicates with the fault prediction module to receive predicted values ​​and prediction residuals, construct a dynamic fault index based on sliding window statistical analysis, and output abnormal early warning signals or normal status signals of the lubrication system. The intelligent control module is connected to the dynamic early warning module and the actuator of the downhole plunger pump lubrication system. It is used to receive early warning signals or normal status signals, and output corresponding lubrication system adjustment commands according to the early warning level to realize adaptive control of lubrication parameters. The data storage module communicates with all the above modules and is used to store the collected raw data, preprocessed data, feature data, prediction results, early warning records and control commands, providing data support for subsequent model optimization and fault tracing.

2. The intelligent early warning system for downhole equipment lubrication based on multi-sensor fault prediction as described in claim 1, characterized in that: The multi-sensor data acquisition module includes a lubrication parameter sensor group, a plunger pump operating parameter sensor group, and a downhole environmental parameter sensor group. The lubrication parameter sensor group includes a lubricating oil temperature sensor, a lubricating oil pressure sensor, a lubricating oil viscosity sensor, a lubricating oil moisture content sensor, and a lubricating oil ferrography sensor, which are used to collect data on the lubricating oil temperature of the downhole plunger pump lubrication system. Lubricating oil pressure Lubricating oil viscosity Lubricating oil moisture content and lubricating oil ferrographic concentration ; The plunger pump operating parameter sensor group includes a pump body vibration sensor, a plunger reciprocating frequency sensor, a pump outlet pressure sensor, and a motor speed sensor, which are used to collect the vibration acceleration of the plunger pump body. Piston reciprocating frequency Pump outlet pressure and motor speed ; The downhole environmental parameter sensor group includes a downhole temperature sensor, a downhole pressure sensor, and a downhole humidity sensor, which are used to collect downhole environmental temperature data. Downhole environmental pressure and underground environmental humidity ; The sampling frequency of all sensors is uniformly set to 5Hz-10Hz. The sampling data is transmitted in real time to the ground data preprocessing module through the downhole wireless transmission module. During the transmission process, the CRC check algorithm is used to verify the data integrity to avoid data loss or transmission errors.

3. The control method of the intelligent early warning system for downhole equipment lubrication, which integrates multi-sensor fault prediction as described in claim 2, is characterized in that: Includes the following steps: S1. Based on the structural characteristics of the downhole plunger pump lubrication system and the high-fault locations, deploy multiple types of sensor groups, set a unified sampling frequency and determine the sampling duration, collect lubrication system parameters, plunger pump operating parameters and downhole environmental parameters in real time, construct a multi-source monitoring raw dataset, use the CRC check algorithm to verify the integrity of data transmission, and complete the preliminary data verification by judging missing values, output the multi-source monitoring raw dataset and sensor anomaly marking information; S2. Receive the original dataset from multi-source monitoring and sensor anomaly labeling information. Based on the multi-parameter correlation characteristics, classify the abnormal data into four categories: shutdown anomaly points, discrete noise points, environmental interference anomaly points, and sensor drift anomaly points. For different types of anomaly points, adopt the corresponding hierarchical cleaning method to complete the abnormal data processing. After eliminating the influence of parameter dimensions by the z-score standardization method, screen the core features with the strongest correlation to lubrication system failures through the improved grey relational analysis method, and output the preprocessed core feature set. S3. Receive the core feature set and reconstruct it into a spatiotemporal input tensor. Use two cascaded convolutional layers to extract multi-scale spatial correlation features. After dimensionality reduction by max pooling, use two LSTM networks to capture the long-term temporal dependence features of the parameters. Introduce a 16-head self-attention mechanism to perform global correlation modeling of the temporal dependence features and complete feature fusion. Output the lubrication state feature vector with fused spatiotemporal features. S4. Select lubricating oil temperature, lubricating oil ferrographic concentration, and lubricating oil viscosity as key fault characterization indicators of the lubrication system. Construct a prediction dataset based on the lubrication state feature vector fused with spatiotemporal features and key fault characterization indicators, and divide it into training set and test set according to the ratio. Construct a prediction model that combines STA feature extraction with fully connected regression. Use the Black Kite algorithm to perform a global optimal search for the model learning rate, batch size, and L2 regularization coefficient. Train the model with the optimal hyperparameters and verify the model performance through multiple indicators. Input the core features after real-time preprocessing into the trained model to predict key indicators in the future and calculate the prediction residuals. Output the predicted values ​​and prediction residuals of the key indicators of the lubrication system. S5. Based on the historical normal operation data of the downhole plunger pump lubrication system, a normal state benchmark model is constructed. The real-time prediction residual is statistically analyzed through a 12-hour sliding window, and the deviation index, fluctuation index, and significance index are calculated. After converting the three types of indices into probability values, they are fused to obtain the dynamic fault index. The adaptive early warning threshold is set and dynamically adjusted in combination with downhole environmental parameters. The corresponding early warning level is output according to the comparison results between the dynamic fault index and the early warning threshold. Based on the early warning level, a hierarchical intelligent control strategy for the lubrication system is formulated and executed. The adjusted lubrication system parameters are collected to complete the control effect feedback and optimization, and the early warning signal and lubrication system control command are output. S6. Retrieve the multi-source monitoring raw data, preprocessed data, feature data, prediction results, early warning records, and control effect data stored in the data storage module. Regularly perform statistical analysis on historical data and identify downhole working conditions where the model prediction deviation exceeds the set value. Supplement the data sample under this working condition and expand the prediction model training set. Retrain and optimize the STA-BKA prediction model using the expanded training set. Adjust the adjustment range of the adaptive early warning threshold and the hierarchical intelligent control strategy in combination with historical early warning records and fault handling results. Update the optimized prediction model, adaptive early warning threshold, and hierarchical intelligent control strategy to the corresponding modules of the system and output the optimized prediction model, adaptive early warning threshold, and hierarchical intelligent control strategy.

4. The control method of the intelligent early warning system for lubrication of downhole equipment integrating multi-sensor fault prediction as described in claim 3, characterized in that: The corresponding layered cleaning method described in step S2 has the following steps: Cleaning for abnormal shutdown points: A cut-off method is used, based on motor speed. Judgment, when Furthermore, if the duration is ≥5 minutes, all lubrication parameter data within that time period will be discarded; For discrete noise point cleaning: the DBSCAN algorithm is adopted, and an optimized adaptive parameter selection strategy is used. Step 1: Calculate the k-nearest neighbor distance for each data point , This represents the total number of valid data points before cleaning. The second step is to plot the k-nearest neighbor distance curve and take the distance corresponding to the inflection point of the curve as the neighborhood radius. ; Step 3: Determine the minimum number of points. : ； in, ； in, To estimate the number of outlier data points, For the first Data points in the domain radius The number of points within; Step 4: Run the DBSCAN algorithm to remove discrete noise points and retain core data points; For cleaning up environmental disturbance anomalies: the kernel density estimation method is used to fit the probability density function of each parameter, and the anomalies corresponding to the density peak value of the probability density function being lower than the set value are removed, thus transforming the multi-peak distribution into a single-peak distribution. For cleaning up sensor drift anomalies: the Sigmoid curve fitting method is used to fit the time series data of each sensor and remove points that deviate from the fitted curve by more than the interquartile range.

5. The control method of the intelligent early warning system for lubrication of downhole equipment integrating multi-sensor fault prediction as described in claim 3, characterized in that: The specific steps in step S2, which involve using an improved grey relational analysis method to screen the core features most strongly correlated with lubrication system failures and outputting the preprocessed core feature set, are as follows: Step 1: Determine the reference sequence as the fault characterization parameters of the lubrication system, denoted as... ; The second step is to determine the comparison sequence as all standardized parameter sequences, denoted as... , Indicates the first Class 1 One sensor Standardized value of time; Step 3: Calculate the grey relational coefficient : ； In the formula, The resolution coefficient, The minimum absolute difference between all sequences. This represents the maximum absolute difference between all sequences. Step 4: Calculate the grey relational degree of each comparison sequence. ,in The length of the time series; Step 5: Set the correlation threshold Filter out The features are used as core features to construct a core feature set. ,in The number of core features.

6. The control method of the intelligent early warning system for lubrication of downhole equipment integrating multi-sensor fault prediction as described in claim 5, characterized in that: The dynamic fault index mentioned in step S5 The expression is as follows: ； in, ； ； ； In the formula, , , These are the deviation probability index, the volatility probability index, and the population stability index, respectively. The cumulative distribution function of the standard normal distribution; It is the deviation index, and , The mean within the window. As the overall deviation benchmark, As a whole discrete reference; The cumulative distribution function is the F-distribution. This represents the number of samples in the real-time window. The number of samples in the test set for the benchmark model; Probability of extreme values; To adjust sensitivity; The adaptive warning threshold expression is as follows: ； ； ； In the formula, for The dynamic threshold for no warning at any given moment; The initial threshold is, and ; , These represent the instantaneous changes in downhole ambient temperature and pressure, respectively. for The dynamic first-level early warning threshold at any given time; Dynamic secondary early warning threshold at any given time; Classification of warning levels: when When there is no warning signal, the lubrication system maintains the current operating parameters and provides feedback on the adjustment effect every 60 minutes. when When the time comes, a first-level warning signal is output, and the lubricating oil circulation flow rate is increased by 10%. The adjustment effect is fed back every 30 minutes. when When the time comes, a level 2 warning signal is output. At this time, the lubricating oil circulation flow rate is increased by 10%, 30% of the lubricating oil is replaced, the plunger pump speed is reduced by 15%, and the adjustment effect is fed back every 15 minutes. when When the time comes, a level three warning signal will be output, the piston pump will stop running, the oil inlet valve of the lubrication system will be closed, and an audible and visual alarm will be issued.

Images