A six-dimensional judgment intelligent operation and maintenance diagnosis method and system for an oilfield gathering station

Abstract

The application discloses a kind of six-dimensional judgment intelligent operation and maintenance diagnosis methods and systems for oilfield gathering and transportation station, belong to oilfield gathering and transportation intelligent operation and maintenance technical field, including data acquisition and single point position model establishment module, six-dimensional analysis and determination module, six-dimensional fusion processing module, feedback updating moduleData acquisition and single point position model establishment module, access oilfield gathering and transportation station each sensor multi-source measuring point data, including oilfield gathering and transportation station flow, pressure, temperature, liquid level, water content, pump frequency / current, valve position / state position, furnace gas quantity / air door, station control alarm position data.The application discloses a kind of six-dimensional judgment intelligent operation and maintenance diagnosis methods and systems for oilfield gathering and transportation station, by integrating multi-link analysis, dynamic threshold adjustment, closed-loop optimization mechanism and safety guarantee design, solve current diagnosis poor stability, insufficient precision, decision support weak and other problems, provide reliable technical support for the efficient and stable operation of oilfield gathering and transportation station.

Description

Technical Field

[0001] This invention belongs to the field of intelligent operation and maintenance technology for oilfield gathering and transportation, specifically a six-dimensional judgment intelligent operation and maintenance diagnosis method and system for oilfield gathering and transportation stations. Background Technology

[0002] As a key hub for oil and gas production, oilfield gathering and transportation stations undertake core tasks such as oil and gas collection, processing, and transportation. The operating status of their equipment directly affects the oilfield's production efficiency and safety. As oilfield development enters the middle and late stages, the equipment at gathering and transportation stations gradually ages, and the operating conditions become increasingly complex. Coupled with the corrosive and high-pressure characteristics of oil and gas media, the risk of equipment failure increases significantly.

[0003] Currently, the operation and maintenance diagnosis of oilfield gathering and transportation stations mostly relies on a single detection link or fixed threshold judgment, lacking multi-dimensional and comprehensive state perception and cross-verification mechanisms. Fault judgment is based only on single mechanism calculations or limited operating data, failing to take into account multiple factors such as design boundaries, process correlations, and big data patterns. This results in narrow diagnostic coverage and insufficient stability. At the same time, the fixed threshold judgment mode is seriously out of touch with actual dynamic operating conditions, making it difficult to adapt to equipment status changes under different production stages and media conditions. This often leads to missed detections of real anomalies or misjudgments of random fluctuations. Furthermore, traditional operation and maintenance diagnosis often only outputs simple anomaly conclusions, lacking confidence analysis of the judgment results, time window positioning, and tracing of related objects. The resulting diagnostic results lack a complete chain of evidence, making it difficult for operation and maintenance personnel to quickly locate the root cause of anomalies. Existing solutions that rely on fixed limits or single models are prone to false alarms, missed alarms, and "operation with defects" under non-steady-state and complex disturbances, directly affecting metering accuracy, process safety, and energy economy. Summary of the Invention

[0004] This invention aims to solve at least one of the technical problems existing in the prior art. To this end, this invention proposes a six-dimensional intelligent operation and maintenance diagnosis method and system for oilfield gathering and transportation stations, which solves the aforementioned technical problems by improving the detection and processing methods.

[0005] To achieve the above objectives, the present invention adopts the following technical solution:

[0006] A six-dimensional intelligent operation and maintenance diagnostic system for oilfield gathering and transportation stations includes a data acquisition and single-point model establishment module, a six-dimensional analysis and judgment module, a six-dimensional fusion processing module, and a feedback and update module.

[0007] The data acquisition and single-point model establishment module accesses multi-source measurement data from various sensors in the oilfield gathering and transportation station, including flow rate, pressure, temperature, liquid level, water content, pump frequency / current, valve position / status position, furnace gas volume / damper, and station control alarm data within the oilfield gathering and transportation station. The module cleans, aligns, and extracts features from the accessed measurement data to construct a computable single-point model.

[0008] The six-dimensional analysis and judgment module, based on the single-point model of the obtained multi-source measurement data, outputs abnormal status and evidence for each point through the six-dimensional analysis link. The six dimensions include design boundary judgment, operation boundary judgment, process correlation verification, mechanism correlation verification, mechanism calculation and deduction, and big data reasoning analysis. Each link outputs a confidence score and evidence string, and the results of each link are summarized into a score vector.

[0009] The six-dimensional fusion processing module fuses the six-dimensional evidence output by the points, generates point credibility scores and anomaly levels, maps the anomaly levels to preset actions based on the action matrix mapping, and sets Boolean logic conditions to ensure control security.

[0010] The feedback update module collects execution results and manually reviewed samples, and writes them back to the model library. It detects changes in model performance based on KL divergence, updates the running boundary and residual threshold using incremental learning, and slowly deploys new parameters through shadow mode and A / B testing. It calculates model health based on false positive rate and false negative rate.

[0011] Furthermore, the data acquisition and single-point model establishment module specifically includes the following:

[0012] Sensor data from oilfield gathering and transportation stations is accessed via industrial protocols. Timestamps at each access point are aligned using the IEEE 1588 precision clock synchronization protocol. Missing timestamps are supplemented using linear interpolation. Based on the nearest valid points , value , Calculation obtained It identifies and removes outliers based on the sliding window Z-score, and simultaneously establishes a local cache to back up data;

[0013] The original signal is converted to engineering units and normalized to 0-100% of the measurement range. Based on the operating parameters, a decision tree binning algorithm is used to generate bin labels, which includes the following steps:

[0014] Historical data, including timestamps and operational events, is collected. A binning algorithm is used to generate bucket labels, and a unique identifier is assigned to each bucket. The information entropy of a single bucket b is calculated. ,in Let k represent the probability distribution of the i-th class of data within a bucket, and k be the number of classes. The goal is to find the optimal bucketing scheme that maximizes the overall entropy reduction, i.e.:

[0015] ;

[0016] in These represent the number of samples in the sub-bucket and the number of samples in the sub-bucket, respectively. Given the number of samples in the parent bucket, the system generates the mean and gradient of sliding window statistical features, quantile features Q1 and Q3, and steady-state labels, providing multi-granular input for six-dimensional decision-making. The sliding window mean is the arithmetic mean of the data points within the window, the gradient is calculated based on the first difference of the time series, and the steady-state label is marked as a steady-state segment when the variance within the sliding window is below a threshold.

[0017] Furthermore, the specific steps for constructing the computable single-point model are as follows:

[0018] Static attributes, dynamic states, upstream and downstream relationships and mechanism parameters are established for each measuring point to form a callable point model. The static attributes are static parameters extracted from design documents or nameplates, including design upper and lower limits and rated values. The attributes are encoded into key-value pair structures.

[0019] Dynamic states include real-time data and historical data sequences. The real-time sequence is the latest sampled value from the access, and the historical data is the cached data within the past n time periods. It also includes steady-state markers and event markers.

[0020] The upstream and downstream relationships are based on the P&ID topology diagram, which assigns upstream and downstream pointers to each UPO. The mechanism parameters are the equipment characteristic parameters.

[0021] Furthermore, the six-dimensional analysis and judgment module analyzes and judges the measurement point data output by the data acquisition and single-point model establishment module based on the six-dimensional link, specifically including the following steps:

[0022] In the design boundary determination, the measured value x of the measuring point data is compared with the nameplate limit. and Perform hard threshold judgment when or If it is, it is marked as an exception, where and These are the upper and lower design limits for the current measurement point in the static attributes.

[0023] During the boundary judgment process, data corresponding to the bucket is extracted from the historical dataset based on the bucket label. Calculate the normal band of dynamic soft boundary [ ],in Given values ​​of 0.05 and 0.95 respectively, calculate the current value x and the average value of the buckets. deviation ,in For the standard deviation of the bins, when Output an over-policy alarm, among which The threshold for boundary determination;

[0024] Process correlation verification requires verifying the consistency of causal chain responses based on upstream and downstream pointers at measurement points, and calculating the actual response time window. That is, the time difference between the change from the upstream point to the change at the local point, which is scored through consistency. Quantifying bias, when Mark suspicious links in time;

[0025] Mechanism correlation verification requires the use of a simplified mechanism model to calculate the expected value. Based on residuals ,in The three sigma criterion based on historical residual distribution in Time marker anomaly;

[0026] Mechanism calculation and deduction need to back-calculate control parameters around production targets, and forward-evaluate the compliance margin and safety margin of process variables, and directly output quantitative parameter adjustment suggestions;

[0027] Big data inference analysis requires calculating the current value using a sliding window based on historical baseline statistics. Abnormal scores ,when Anomalies are flagged and trends are recorded. At the end of each update cycle, new data is added to the historical dataset and recalculated. ,in This is the threshold for inference analysis.

[0028] Furthermore, each link outputs a confidence score and an evidence string, and the results of each link are aggregated into a score vector, specifically including the following steps:

[0029] For the judgment results of the six-dimensional link output, assign a continuous confidence score. ,in The range is between 0 and 1, and it is used for design boundary determination. or hour ,on the contrary ;

[0030] In operational boundary determination ,in For the current deviation, This represents the largest deviation in history.

[0031] In process correlation verification In mechanism correlation verification where e is the residual For the residual standard deviation, the mechanism is calculated and deduced. , of which Confidence level of inverse calculation error and Safety margin confidence, of which , , These are the inverse calculation error and the error tolerance, respectively. These represent the safety margin and the safety margin reference value, respectively, and the confidence level for big data inference analysis. , These represent the anomaly score and the maximum historical anomaly score, respectively. Key values ​​of the measurement points, time windows, associated objects, and alarm numbers are encapsulated into an evidence string. The confidence scores of the six links are combined into a six-dimensional scoring vector. .

[0032] Furthermore, the six-dimensional fusion processing module fuses the six-dimensional evidence output by the points, generates point credibility scores and anomaly levels, maps the anomaly levels to preset actions based on the action matrix mapping, and sets Boolean logic conditions to ensure control safety. Specifically, it includes the following steps:

[0033] The scores and evidence strings of the six dimensions are summarized, and the weights of the confidence scores of the six dimensions are set according to the priority to calculate the comprehensive confidence score. The confidence interval is calculated and the evidence conflict is handled to generate the point confidence score and anomaly level.

[0034] The exception level is mapped to a preset action, and a safety interlock is implemented based on Boolean logic conditions to output the processing action.

[0035] Furthermore, the aggregated six-dimensional output scores and evidence strings are weighted according to priority allocation to calculate the comprehensive confidence score, calculate the confidence interval, handle evidence conflicts, and generate point confidence scores and anomaly levels. The specific steps are as follows:

[0036] Weights are assigned to each dimension based on the reliability of the evidence. ,satisfy Calculate the overall confidence level ,in Score the confidence level;

[0037] Calculate the mean and standard deviation of the six-dimensional rating vector S to set the confidence interval L, where:

[0038] ;

[0039] in, represent the mean, standard deviation, and sample size of the six-dimensional scoring vector S, respectively. This corresponds to a 95% confidence level. Output uncertainty index ;

[0040] Through U pair Make corrections, including correcting the confidence level. ,when When the evidence is consistent, use it directly. As a basis for credibility, when This triggers conflict handling, including priority arbitration and uncertainty-driven degradation. Priority arbitration directly marks anomalies as abnormal when the design boundary dimension is determined to be anomaly. Uncertainty-driven degradation, on the other hand, involves... At that time, the credibility score anomaly level will be forcibly downgraded and labeled "high uncertainty". This is the uncertainty threshold;

[0041] Calculate credibility score based on modified confidence level ,in These represent the extreme values ​​of the historical revised confidence level, when ≥0.8 and U< When the abnormal level is judged as normal, ≥0.8 and U≥ Or 0.5≤ When <0.8, the anomaly level is judged as suspicious. If the value is less than 0.5, the anomaly level is determined to be abnormal, and a conflict resolution record field is added to the evidence string.

[0042] Furthermore, the process of mapping exception levels to preset actions, implementing safety interlocks based on Boolean logic conditions, and outputting processing actions specifically includes the following steps:

[0043] For each determined anomaly level, an anomaly level-action mapping matrix is ​​constructed, showing the mapping relationships. , where K is a two-dimensional matrix, Level is the anomaly level, and Ac is the pre-defined action for the anomaly;

[0044] At the same time, Boolean logic conditions are set to ensure that control actions are performed only under safe conditions, i.e. , These represent safety condition 1, safety condition 2, and hazardous condition, respectively. The data collection and execution receipts are used to calculate the achievement rate. When the achievement rate is below 0.9, an alternative strategy is triggered. These represent the numerical values ​​of the actual achieved result or state, and the preset target value, respectively.

[0045] Furthermore, the feedback update module collects execution results and manually reviewed samples, and writes them back to the model library. It detects model performance changes based on KL divergence, updates the running boundary and residual threshold using incremental learning, and slowly deploys new parameters through shadow mode and A / B testing. It calculates model health based on false positive rate and false negative rate. The specific steps are as follows:

[0046] Collect the execution results of the six-dimensional fusion processing module, manually reviewed samples, and new system operation data. Classify the data by type and write it back to the corresponding model library. The model library includes a success case library, a false positive sample library, a false negative sample library, and a basic historical database. Based on a fixed-period comparison of the model's distribution on the old and new data, where the old distribution P represents the historical data distribution on which the model was trained, and the new distribution Q represents the distribution of recently incoming online data, calculate the KL divergence:

[0047] ;

[0048] Where x is the value of the random variable, when > When a significant model drift is detected, model reconstruction is triggered. The divergence threshold;

[0049] For operational boundaries and residual thresholds, an exponentially weighted moving average is used for online smooth updates. New parameters are slowly deployed using shadow mode and A / B testing. A model health score is calculated weekly. ,in The system identifies false alarms, missed alarms, and the total number of events. An alarm is triggered when H is below 0.8.

[0050] A six-dimensional intelligent operation and maintenance diagnosis method for oilfield gathering and transportation stations includes the following steps:

[0051] S1. Access multi-source measurement data from various sensors in the oilfield gathering and transportation station, including flow rate, pressure, temperature, liquid level, water content, pump frequency / current, valve position / status position, furnace gas volume / damper, and station control alarm data. Clean, align, and extract features from the accessed measurement data to construct a computable single-point model.

[0052] S2. Based on the obtained multi-source measurement data single-point model, the abnormal status and evidence are output for each point through a six-dimensional analysis link. The six dimensions include design boundary judgment, operation boundary judgment, process correlation verification, mechanism correlation verification, mechanism calculation and deduction, and big data reasoning analysis. Each link outputs a confidence score and evidence string, and the results of each link are summarized into a score vector.

[0053] S3. Integrate the six-dimensional evidence output from the points, generate point credibility scores and anomaly levels, map the anomaly levels to preset actions based on the action matrix mapping, set Boolean logic conditions to ensure control safety, and output processing actions.

[0054] S4. Collect execution results, manually review samples, and write them back to the model library. Detect model performance changes based on KL divergence, use incremental learning to update the running boundary and residual threshold, and slowly deploy new parameters through shadow mode and A / B testing. Calculate model health based on false positive rate and false negative rate.

[0055] Compared with the prior art, the beneficial effects of the present invention are:

[0056] 1. In this invention, by running six links in parallel at the measurement points—design boundary, operation boundary, process correlation, mechanism correlation, mechanism calculation, and big data inference—mutual verification is achieved, significantly improving the stability and coverage of identification. Soft boundaries are dynamically generated according to operating parameters and slowly put online for self-calibration, solving the problem of the disconnect between fixed thresholds and actual operating conditions, and outputting parameter adjustment suggestions that are closer to the field.

[0057] 2. In this invention, through parallel analysis of six dimensions—design boundary, operational boundary, process correlation, mechanism correlation, mechanism calculation, and big data reasoning—the system cross-validates the equipment status from multiple independent perspectives, avoiding the limitations of a single detection method. It can effectively distinguish between real anomalies and random fluctuations, thereby improving diagnostic accuracy and reducing false alarm and false negative rates. It not only outputs anomaly conclusions but also generates an evidence string containing confidence scores, time windows, and related objects for each judgment result to form a clear chain of evidence, enabling maintenance personnel to quickly understand the root cause of the anomaly.

[0058] 3. In this invention, the control commands are driven by the fusion processing module, and the execution effect and human experience are written back to the model library by the feedback module. This enables the automatic execution of the handling actions and the self-correction of the model parameters, allowing the system to continuously evolve and adapt to changes in equipment operating conditions and maintain high reliability over a long period of time. At the data acquisition layer, local caching and breakpoint resume transmission ensure that data is not lost when the network is interrupted. At the control layer, the safety of the automatic control commands is ensured by safety interlock logic and manual confirmation mechanism. Attached Figure Description

[0059] Figure 1 This is a block diagram of a six-dimensional intelligent operation and maintenance diagnostic system for oilfield gathering and transportation stations according to the present invention. Detailed Implementation

[0060] The technical solution of the present invention will be clearly and completely described below with reference to the embodiments. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0061] Example 1:

[0062] like Figure 1 As shown, a six-dimensional intelligent operation and maintenance diagnostic system for oilfield gathering and transportation stations includes a data acquisition and single-point model establishment module, a six-dimensional analysis and judgment module, a six-dimensional fusion processing module, and a feedback update module.

[0063] The data acquisition and single-point model building module integrates multi-source measurement data from various sensors in the oilfield gathering and transportation station, including data on flow rate, pressure, temperature, liquid level, water content, pump frequency / current, valve position / status, furnace gas volume / damper volume, and station control alarm status. It cleans, aligns, and extracts features from the integrated measurement data to construct a computable single-point model, specifically including the following:

[0064] Sensor data from oilfield gathering and transportation stations is accessed via industrial protocols. Timestamps at each access point are aligned using the IEEE 1588 precision clock synchronization protocol. Missing timestamps are supplemented using linear interpolation. Based on the nearest valid points , value , Calculation obtained It identifies and removes outliers based on the sliding window Z-score, and simultaneously establishes a local cache to back up data;

[0065] It should be noted that industrial protocols include OPC UA, Modbus, and Profibus, which integrate multi-source sensor data such as flow rate, pressure, temperature, liquid level, and moisture content. They support interface types such as 4-20mA, HART, and DI / DO, and are compatible with existing systems like SCADA, achieving unified integration of heterogeneous protocols. When identifying glitch data based on a sliding window Z-score, the data points are... ,calculate ,in These represent the window mean and window standard deviation, respectively. If an outlier is detected, it will be considered an anomaly and removed. The local cache is a ring buffer deployed on the edge side, which supports data caching for more than 72 hours when the network is interrupted. The cached data is indexed by timestamp, and the interrupted data is resumed after the connection is restored to ensure data integrity and provide consistent and clean time-series data input for subsequent link judgment.

[0066] The original signal is converted to engineering units and normalized to 0-100% of the measurement range. Based on the operating parameters, a decision tree binning algorithm is used to generate bin labels, which includes the following steps:

[0067] Historical data, including timestamps and operational events, is collected. A binning algorithm is used to generate bucket labels, and a unique identifier is assigned to each bucket. The information entropy of a single bucket b is calculated. ,in Let k represent the probability distribution of the i-th class of data within a bucket, and k be the number of classes. The goal is to find the optimal bucketing scheme that maximizes the overall entropy reduction, i.e.:

[0068] ;

[0069] in These represent the number of samples in the sub-bucket and the number of samples in the sub-bucket, respectively. Given the number of samples in the parent bucket, the system generates the mean and gradient of sliding window statistical features, quantile features Q1 and Q3, and steady-state labels, providing multi-granular input for six-dimensional decision-making. The sliding window mean is the arithmetic mean of the data points within the window, the gradient is calculated based on the first difference of the time series, and the steady-state label is marked as a steady-state segment when the variance within the sliding window is below a threshold.

[0070] It should be noted that the operating parameters include contextual parameters such as season, load, pump configuration, and hot washing / water mixing status. A lower entropy value indicates higher data purity within the bucket. Optimal bucketing is achieved by recursively selecting the split point with the largest entropy increase. The algorithm starts from the root node, selecting the optimal split point, for example, using a load value of 50% as the boundary, to divide the data into subsets. This process is repeated until stopping conditions are met, including when the number of samples in the bucket is below a threshold or the entropy reduction is not significant. After collecting the operating parameters, discrete parameters need to be converted into numerical feature vectors. For example, seasonal parameters can be encoded as numerical or unique heat vectors, and load parameters can be normalized to the 0-1 range. A unique label is assigned to each bucket, such as "High Load - Summer - Pump A - No Hot Washing". The assigned label is used for subsequent dynamic boundary calculations. The threshold needs to be set empirically based on actual conditions. Simultaneously, based on the P&ID process topology and pipeline volume parameters, the upstream and downstream data propagation delays are calculated, and the sequences are aligned to eliminate causal misalignment caused by transmission delays. The time delay is... V and Q represent the pipe volume and flow rate, respectively. Due to the inherent delay of the sensor, the parent node is the node or data subset currently being segmented during the binning process, and H (parent node) is the entropy of the parent node's dataset.

[0071] The specific steps for constructing a computable single-point model are as follows:

[0072] Static attributes, dynamic states, upstream and downstream relationships and mechanism parameters are established for each measuring point to form a callable point model. The static attributes are static parameters extracted from design documents or nameplates, including design upper and lower limits and rated values. The attributes are encoded into key-value pair structures.

[0073] Dynamic states include real-time data and historical data sequences. The real-time sequence is the latest sampled value from the access, and the historical data is the cached data within the past n time periods. It also includes steady-state markers and event markers.

[0074] It should be noted that n is usually set to 24 hours, but can also be adjusted according to the time situation. The steady state marker needs to be determined by the variance threshold. Event markers include pump start-up and shutdown events, design upper and lower limits such as pressure range, and the rated value is the set rated value of the equipment.

[0075] The upstream and downstream relationships are based on the P&ID topology diagram, which assigns upstream and downstream pointers to each UPO. The mechanism parameters are the equipment characteristic parameters.

[0076] It should be noted that pointers include correlation types, such as causal and parallel relationships, as well as response time window parameters such as flow rate change propagation delay. For example, the upstream pointer of the pump outlet pressure point points to the pump frequency point, and the downstream pointer points to the valve opening point. Equipment characteristic parameters include, for example, pipeline resistance coefficient, pump efficiency curve, heat exchanger heat transfer coefficient, etc., which can be obtained by consulting the equipment design manual.

[0077] The six-dimensional analysis and judgment module, based on the obtained multi-source measurement data and single-point model, outputs abnormal states and evidence for each point through a six-dimensional analysis link. The six dimensions include design boundary judgment, operational boundary judgment, process correlation verification, mechanism correlation verification, mechanism calculation and deduction, and big data inference analysis. Each link outputs a confidence score and an evidence string. The results of each link are summarized into a score vector, specifically including the following steps:

[0078] In the design boundary determination, the measured value x of the measuring point data is compared with the nameplate limit. and Perform hard threshold judgment when or If it is, it is marked as an exception, where and These are the upper and lower design limits for the current measurement point in the static attributes.

[0079] During the boundary judgment process, data corresponding to the bucket is extracted from the historical dataset based on the bucket label. Calculate the normal band of dynamic soft boundary [ ],in Given values ​​of 0.05 and 0.95 respectively, calculate the current value x and the average value of the buckets. deviation ,in For the standard deviation of the bins, when Output an over-policy alarm, among which The threshold for boundary determination;

[0080] It should be noted that, This indicates the data from the corresponding bucket. The lower quantile calculated in the middle, that is, the α×100% quantile, is the 5th quantile. This represents the 95th quantile. Two quantiles are used to determine the dynamic soft boundary to judge whether the current value x is within the normal operating range. The boundary judgment threshold requires the collection of a large amount of historical operating data, the calculation of the distribution of deviations under normal operating conditions, the calculation of the mean and standard deviation of deviations in historical data, and the setting of the three sigma criterion.

[0081] Process correlation verification requires verifying the consistency of causal chain responses based on upstream and downstream pointers at measurement points, and calculating the actual response time window. That is, the time difference between the change from the upstream point to the change at the local point, which is scored through consistency. Quantifying bias, when Mark suspicious links in time;

[0082] It should be noted that, This is the threshold for process correlation verification. It is necessary to calibrate based on P&ID topology and volume parameters, and to check whether upstream changes produce reasonable responses at this point and downstream in accordance with the time window by combining topology and causal relationship, so as to identify hidden problems such as jamming, leakage, backflow, etc.

[0083] Mechanism correlation verification requires the use of a simplified mechanism model to calculate the expected value. Based on residuals ,in The three sigma criterion based on historical residual distribution in Time marker anomaly;

[0084] It should be noted that simplified mechanism models, such as those for pumps / pipelines, heat exchangers, and tank conservation, need to be obtained based on measurement points. The expected values ​​calculated by the simplified mechanism model are compared with the actual measurements to obtain the residuals, thereby determining physical inconsistencies and parameter drift.

[0085] Mechanism calculation and deduction need to back-calculate control parameters around production targets, and forward-evaluate the compliance margin and safety margin of process variables, and directly output quantitative parameter adjustment suggestions;

[0086] It should be noted that, taking the production target as the set flow rate as an example, by back-calculating control parameters such as pump frequency and valve position, and by forward-calculating process variables such as pressure drop and temperature rise, the margin of compliance is evaluated, i.e., the deviation between the actual value and the target and the distance from the design boundary. Quantitative parameter adjustment suggestions are directly output, such as "increase pump frequency by 5Hz". By back-calculating / forward-calculating around the target and constraints, quantitative adjustment suggestions for pump frequency, valve position, heating power, etc., are directly given. , These represent the pump characteristic coefficient and target flow rate, respectively, and the temperature rise. ,in These represent heating power, density, and specific heat capacity, respectively.

[0087] Big data inference analysis requires calculating the current value using a sliding window based on historical baseline statistics. Abnormal scores ,when Anomalies are flagged and trends are recorded. At the end of each update cycle, new data is added to the historical dataset and recalculated. ,in This is the threshold for inference analysis.

[0088] It should be noted that the inference analysis threshold Specific settings are required based on different non-steady-state anomalies. Different non-steady-state anomaly inference and analysis thresholds are used to identify non-steady-state anomalies such as slow drift, fluctuation amplification, and time delay changes. For example, the slow drift anomaly inference and analysis threshold needs to take into account the rate of data change and the time scale. It is determined based on the range of slow changes under normal conditions in historical data using empirical methods. The update cycle can be set to annually, quarterly, or monthly, depending on the actual application needs.

[0089] Each link outputs a confidence score and an evidence string. The results of each link are then aggregated into a score vector, which includes the following steps:

[0090] For the judgment results of the six-dimensional link output, assign a continuous confidence score. ,in The range is between 0 and 1, and it is used for design boundary determination. or hour ,on the contrary ;

[0091] In operational boundary determination ,in For the current deviation, This represents the largest deviation in history.

[0092] In process correlation verification In mechanism correlation verification where e is the residual For the residual standard deviation, the mechanism is calculated and deduced. , of which Confidence level of inverse calculation error and Safety margin confidence, of which , , These are the inverse calculation error and the error tolerance, respectively. These represent the safety margin and the safety margin reference value, respectively, and the confidence level for big data inference analysis. , These represent the anomaly score and the maximum historical anomaly score, respectively. Key values ​​of the measurement points, time windows, associated objects, and alarm numbers are encapsulated into an evidence string. The confidence scores of the six links are combined into a six-dimensional scoring vector. .

[0093] It should be noted that the error tolerance needs to be set based on historical performance or engineering experience. This is a normal and acceptable safety margin reference value, with inverse calculation error. Where y represents the key parameters that need to be verified, such as pump frequency and outlet pressure. These are the actual measured values ​​of the key parameters. These are the expected values ​​of key parameters, which are theoretically calculated values ​​obtained through back-calculation and forward extrapolation. Key values ​​include measured values, expected values, deviations, residuals, etc. The time window is the data time range on which the judgment is based. The associated object is the ID of the upstream and downstream related points, such as the ID of the upstream pump frequency point that caused the anomaly at this point. The alarm number is the unique identifier of this judgment event. The six judgment links run in independent threads or processes. The scoring vector S is a point in six-dimensional space, and its coordinate values ​​reflect the health of each dimension. The evidence string is an interpretable supplement to the vector S, forming a "data-conclusion-evidence" triple.

[0094] Example 2:

[0095] The six-dimensional fusion processing module fuses the six-dimensional evidence output from the points, generates point credibility scores and anomaly levels, maps the anomaly levels to preset actions based on the action matrix mapping, and sets Boolean logic conditions to ensure control safety. Specifically, it includes the following steps:

[0096] The process involves summarizing the scores and evidence strings from the six dimensions, assigning weights to the confidence levels of each of the six dimensions based on priority, calculating the overall confidence level, calculating the confidence interval, handling evidence conflicts, and generating point confidence scores and anomaly levels. The specific steps are as follows:

[0097] Weights are assigned to each dimension based on the reliability of the evidence. ,satisfy Calculate the overall confidence level ,in Score the confidence level;

[0098] It should be noted that during the weight allocation process, the design boundary must be given the highest weight because it relates to intrinsic safety. Other weights are dynamically adjusted based on historical accuracy. ,in The base weights are 'a' and the learning rate is 'a'.

[0099] Calculate the mean and standard deviation of the six-dimensional rating vector S to set the confidence interval L, where:

[0100] ;

[0101] in, represent the mean, standard deviation, and sample size of the six-dimensional scoring vector S, respectively. This corresponds to a 95% confidence level. Output uncertainty index ;

[0102] It should be noted that the uncertainty index is calculated based on the width of the confidence interval, which is related to the dispersion of the six-dimensional scoring vector data. When the data is more dispersed, the confidence interval will be wider and the uncertainty index will increase, while the uncertainty index will decrease when the data is more concentrated.

[0103] Through U pair Make corrections, including correcting the confidence level. ,when When the evidence is consistent, use it directly. As a basis for credibility, when This triggers conflict handling, including priority arbitration and uncertainty-driven degradation. Priority arbitration directly marks anomalies as abnormal when the design boundary dimension is determined to be anomaly. Uncertainty-driven degradation, on the other hand, involves... At that time, the credibility score anomaly level will be forcibly downgraded and labeled "high uncertainty". This is the uncertainty threshold;

[0104] It should be noted that forcibly downgrading the credibility score, for example, downgrading the abnormality level to suspicious, This is the uncertainty threshold, which can be set to 0.1 or adjusted according to the actual situation.

[0105] Calculate credibility score based on modified confidence level ,in These represent the extreme values ​​of the historical revised confidence level, when ≥0.8 and U< When the abnormal level is judged as normal, ≥0.8 and U≥ Or 0.5≤ When <0.8, the anomaly level is judged as suspicious. If the value is less than 0.5, the anomaly level is determined to be abnormal, and a conflict resolution record field is added to the evidence string.

[0106] It should be noted that the new conflict resolution record field in the evidence string includes the uncertainty index U value, and a manual review suggestion mark is added under the "high uncertainty" label.

[0107] The exception level is mapped to a preset action, a safety interlock is implemented based on Boolean logic conditions, and the processing action is output. The specific steps include:

[0108] For each determined anomaly level, an anomaly level-action mapping matrix is ​​constructed, showing the mapping relationships. , where K is a two-dimensional matrix, Level is the anomaly level, and Ac is the pre-defined action for the anomaly;

[0109] It should be noted that after constructing the exception level-action mapping matrix, it is also necessary to determine the action priority. This ensures that actions with high risk levels are prioritized. Risk level indicates the degree of potential risk brought about by an anomaly, and is usually assessed based on factors such as the nature of the anomaly, its scope of impact, and the potential losses. For example, in a power system, a short-circuit fault in equipment may be assessed as a high-risk level, while some minor signal fluctuations may be assessed as a low-risk level. K is a two-dimensional matrix, where each row corresponds to an anomaly level and each column corresponds to a possible action. Matrix elements represent the specific actions to be taken under that anomaly level, serving as a mapping relationship between anomaly levels and specific actions. By querying this matrix, the actions to be taken under different anomaly levels can be quickly determined.

[0110] At the same time, Boolean logic conditions are set to ensure that control actions are performed only under safe conditions, i.e. , These represent safety condition 1, safety condition 2, and hazardous condition, respectively. The data collection and execution receipts are used to calculate the achievement rate. When the achievement rate is below 0.9, an alternative strategy is triggered. These represent the numerical values ​​of the actual achieved result or state, and the preset target value, respectively.

[0111] It should be noted that safety condition 1 and safety condition 2 are independent, but together they constitute the safety prerequisites for allowing actions to be performed. In specific application scenarios, a safety condition can be a specific safety requirement or safety state that the system meets. A dangerous condition represents a state in the system that may lead to a dangerous situation. For high-risk actions such as full-site load adjustment, manual confirmation is required, with a timeout period set. If confirmation is not received within the timeout period, an escalation notification is issued, typically set to 30 seconds. Additionally, a local degradation strategy is enabled during weak network / network outages, prioritizing the retention of design boundary judgment functions. Alternative strategies can collect relevant data from past system executions of various strategies, including data under different strategies. And V, through data analysis, identify strategies with high effectiveness and their applicable conditions, map the action matrix according to the anomaly level, and after safety interlocking and necessary manual confirmation, issue them to DCS / PLC / intelligent actuators, such as valve position adjustment, frequency conversion load reduction, pump group switching, heating start and stop, etc.

[0112] The feedback update module collects execution results and manually reviewed samples, and writes them back to the model library. It detects model performance changes based on KL divergence, updates the runtime boundary and residual threshold using incremental learning, and slowly deploys new parameters through shadow mode and A / B testing. It calculates model health based on false positive and false negative rates. The specific steps are as follows:

[0113] Collect the execution results of the six-dimensional fusion processing module, manually reviewed samples, and new system operation data. Classify the data by type and write it back to the corresponding model library. The model library includes a success case library, a false positive sample library, a false negative sample library, and a basic historical database. Based on a fixed-period comparison of the model's distribution on the old and new data, where the old distribution P represents the historical data distribution on which the model was trained, and the new distribution Q represents the distribution of recently incoming online data, calculate the KL divergence:

[0114] ;

[0115] Where x is the value of the random variable, when > When a significant model drift is detected, model reconstruction is triggered. The divergence threshold;

[0116] It should be noted that, It is usually set to 0.1. It is non-negative. The larger the value, the greater the difference between distribution Q and P. When it is greater than the threshold, it means that the current model can no longer describe the recent working conditions well and needs to be adjusted. The fixed period is usually set to 24 hours.

[0117] For operational boundaries and residual thresholds, an exponentially weighted moving average is used for online smooth updates. New parameters are slowly deployed using shadow mode and A / B testing. A model health score is calculated weekly. ,in The system identifies false alarms, missed alarms, and the total number of events. An alarm is triggered when H is below 0.8.

[0118] It should be noted that after each diagnosis, the entire chain of evidence for a single diagnosis is visualized in the form of a knowledge graph. Nodes represent data points, alarm events, and handling actions, while edges represent causal relationships and judgment logic flows. When a new model or threshold needs to be deployed, it is first run in "shadow mode" version A on some locations or traffic flows, in parallel with the old version B. During the preset testing period, indicators such as the false alarm rate of the two versions are collected. Hypothesis testing is used to calculate that if the result is less than the significance level, such as 0.05, then a full switch to the new version is made. The specific steps for the exponentially weighted moving average are as follows: for the new threshold... ,in These represent the smoothing factor and the statistic calculated based on the new sample, respectively.

[0119] Example 3:

[0120] A six-dimensional intelligent operation and maintenance diagnosis method for oilfield gathering and transportation stations includes the following steps:

[0121] S1. Access multi-source measurement data from various sensors in the oilfield gathering and transportation station, including flow rate, pressure, temperature, liquid level, water content, pump frequency / current, valve position / status position, furnace gas volume / damper, and station control alarm data. Clean, align, and extract features from the accessed measurement data to construct a computable single-point model.

[0122] S2. Based on the obtained multi-source measurement data single-point model, the abnormal status and evidence are output for each point through a six-dimensional analysis link. The six dimensions include design boundary judgment, operation boundary judgment, process correlation verification, mechanism correlation verification, mechanism calculation and deduction, and big data reasoning analysis. Each link outputs a confidence score and evidence string, and the results of each link are summarized into a score vector.

[0123] S3. Integrate the six-dimensional evidence output from the points, generate point credibility scores and anomaly levels, map the anomaly levels to preset actions based on the action matrix mapping, set Boolean logic conditions to ensure control safety, and output processing actions.

[0124] S4. Collect execution results, manually review samples, and write them back to the model library. Detect model performance changes based on KL divergence, use incremental learning to update the running boundary and residual threshold, and slowly deploy new parameters through shadow mode and A / B testing. Calculate model health based on false positive rate and false negative rate.

[0125] This invention presents a six-dimensional intelligent operation and maintenance diagnostic method and system for oilfield gathering and transportation stations. During operation, it utilizes six parallel chains—design boundary, operational boundary, process correlation, mechanism correlation, mechanism calculation, and big data inference—running at measurement points to mutually verify each other, significantly improving identification stability and coverage. It dynamically generates soft boundaries based on operating parameters and performs slow online self-calibration, resolving the disconnect between fixed thresholds and actual operating conditions, and outputting parameter adjustment suggestions closer to the field. Through parallel analysis of these six dimensions—design boundary, operational boundary, process correlation, mechanism correlation, mechanism calculation, and big data inference—the system cross-verifies equipment status from multiple independent perspectives, avoiding the limitations of single detection methods and effectively distinguishing between real anomalies and random fluctuations. This improves diagnostic accuracy and reduces false alarm and false negative rates. It not only outputs anomaly conclusions but also generates an evidence string for each judgment result, including confidence scores, time windows, and related objects, forming a clear chain of evidence. This allows maintenance personnel to quickly understand the root cause of the anomaly. The fusion processing module drives control commands, and the feedback module writes the execution results and human experience back to the model library, enabling automatic execution of actions and self-correction of model parameters. This allows the system to continuously evolve and adapt to changes in equipment operating conditions, maintaining high reliability over the long term. At the data acquisition layer, local caching and breakpoint resumption ensure no data loss during network interruptions. At the control layer, security interlock logic and manual confirmation mechanisms ensure the safety of automatic control commands.

[0126] In the embodiments provided by this invention, it should be understood that the disclosed devices, apparatuses, and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of modules is only a logical functional division, and other division methods may exist in actual implementation. Modules described as separate components may or may not be physically separated, and components shown as modules may or may not be physical units, i.e., they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of the method in this embodiment according to actual needs.

[0127] The above embodiments are only used to illustrate the technical methods of the present invention and are not intended to limit it. 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 be made to the technical methods of the present invention without departing from the spirit and scope of the technical methods of the present invention.

Claims

1. A six-dimensional intelligent operation and maintenance diagnostic system for oilfield gathering and transportation stations, characterized in that: It includes a data acquisition and single-point model building module, a six-dimensional analysis and judgment module, a six-dimensional fusion processing module, and a feedback update module; The data acquisition and single-point model establishment module accesses multi-source measurement data from various sensors in the oilfield gathering and transportation station, including flow rate, pressure, temperature, liquid level, water content, pump frequency / current, valve position / status position, furnace gas volume / damper, and station control alarm data within the oilfield gathering and transportation station. The module cleans, aligns, and extracts features from the accessed measurement data to construct a computable single-point model. The six-dimensional analysis and judgment module, based on the single-point model of the obtained multi-source measurement data, outputs abnormal status and evidence for each point through the six-dimensional analysis link. The six dimensions include design boundary judgment, operation boundary judgment, process correlation verification, mechanism correlation verification, mechanism calculation and deduction, and big data reasoning analysis. Each link outputs a confidence score and evidence string, and the results of each link are summarized into a score vector. In the design boundary determination, the measured value x of the measurement point data is compared with the nameplate limit. and Perform hard threshold judgment when or If it is, it is marked as an exception, where and These are the upper and lower design limits for the current measurement point in the static attributes. During the boundary judgment process, data corresponding to the bucket is extracted from the historical dataset based on the bucket label. Calculate the normal band of dynamic soft boundary [ ],in Given values ​​of 0.05 and 0.95 respectively, calculate the current value x and the average value of the buckets. deviation ,in For the standard deviation of the bins, when Output an over-policy alarm, among which The threshold for boundary determination; Process correlation verification requires verifying the consistency of causal chain responses based on upstream and downstream pointers at measurement points, and calculating the actual response time window. That is, the time difference between the change from the upstream point to the change at the local point, which is scored through consistency. Quantifying bias, when Mark suspicious links in time; Mechanism correlation verification requires the use of a simplified mechanism model to calculate the expected value. Based on residuals ,in The three sigma criterion based on historical residual distribution in Time marker exception, where The standard deviation of the residuals; Mechanism calculation and deduction need to back-calculate control parameters around production targets, and forward-evaluate the compliance margin and safety margin of process variables, and directly output quantitative parameter adjustment suggestions; Big data inference analysis requires calculating the current value using a sliding window based on historical baseline statistics. Abnormal scores ,when Anomalies are flagged and trends are recorded. At the end of each update cycle, new data is added to the historical dataset and recalculated. ,in Threshold for inference analysis; The six-dimensional fusion processing module fuses the six-dimensional evidence output by the points, generates point credibility scores and anomaly levels, maps the anomaly levels to preset actions based on the action matrix mapping, and sets Boolean logic conditions to ensure control security. The feedback update module collects execution results and manually reviewed samples, and writes them back to the model library. It detects changes in model performance based on KL divergence, updates the running boundary and residual threshold using incremental learning, and slowly deploys new parameters through shadow mode and A / B testing. It calculates model health based on false positive rate and false negative rate.

2. The six-dimensional intelligent operation and maintenance diagnostic system for oilfield gathering and transportation stations according to claim 1, characterized in that: The data acquisition and single-point model establishment module specifically includes the following: Sensor data from oilfield gathering and transportation stations is accessed via industrial protocols. Timestamps at each access point are aligned using the IEEE 1588 precision clock synchronization protocol. Missing timestamps are supplemented using linear interpolation. Based on the nearest valid points , value , Calculation obtained It identifies and removes outliers based on the sliding window Z-score, and simultaneously establishes a local cache to back up data; The original signal is converted to engineering units and normalized to 0-100% of the measurement range. Based on the operating parameters, a decision tree binning algorithm is used to generate bin labels, which includes the following steps: Historical data, including timestamps and operational events, is collected. A binning algorithm is used to generate bucket labels, and a unique identifier is assigned to each bucket. The information entropy of a single bucket b is calculated. ,in Let k represent the probability distribution of the i-th class of data within a bucket, and k be the number of classes. The goal is to find the optimal bucketing scheme that maximizes the overall entropy reduction, i.e.: in These represent the number of samples in the sub-bucket and the number of samples in the sub-bucket, respectively. Given the number of samples in the parent bucket, the system generates the mean and gradient of sliding window statistical features, quantile features Q1 and Q3, and steady-state labels, providing multi-granular input for six-dimensional decision-making. The sliding window mean is the arithmetic mean of the data points within the window, the gradient is calculated based on the first difference of the time series, and the steady-state label is marked as a steady-state segment when the variance within the sliding window is below a threshold.

3. The six-dimensional intelligent operation and maintenance diagnostic system for oilfield gathering and transportation stations according to claim 2, characterized in that: The specific steps for constructing a computable single-point model are as follows: Static attributes, dynamic states, upstream and downstream relationships and mechanism parameters are established for each measuring point to form a callable point model. The static attributes are static parameters extracted from design documents or nameplates, including design upper and lower limits and rated values. The attributes are encoded into key-value pair structures. Dynamic states include real-time data and historical data sequences. The real-time sequence is the latest sampled value from the access, and the historical data is the cached data within the past n time periods. It also includes steady-state markers and event markers. The upstream and downstream relationships are based on the P&ID topology diagram, which assigns upstream and downstream pointers to each UPO. The mechanism parameters are the equipment characteristic parameters.

4. The six-dimensional intelligent operation and maintenance diagnostic system for oilfield gathering and transportation stations according to claim 3, characterized in that: Each link outputs a confidence score and an evidence string. The results of each link are then aggregated into a score vector, which includes the following steps: For the judgment results of the six-dimensional link output, assign a continuous confidence score. ,in The range is between 0 and 1, and it is used for design boundary determination. or hour ,on the contrary ; In operational boundary determination ,in For the current deviation, This represents the largest deviation in history. In process correlation verification In mechanism correlation verification Where e is the residual, For the residual standard deviation, the mechanism is calculated and deduced. ,in The confidence level of the inverse calculation error, For safety margin confidence, where , , These are the inverse calculation error and the error tolerance, respectively. These represent the safety margin and the safety margin reference value, respectively, and the confidence level for big data inference analysis. , These represent the anomaly score and the maximum historical anomaly score, respectively. Key values ​​of the measurement points, time windows, associated objects, and alarm numbers are encapsulated into an evidence string. The confidence scores of the six links are combined into a six-dimensional scoring vector. .

5. The six-dimensional intelligent operation and maintenance diagnostic system for oilfield gathering and transportation stations according to claim 4, characterized in that: The six-dimensional fusion processing module fuses the six-dimensional evidence output by the points, generates point credibility scores and anomaly levels, maps the anomaly levels to preset actions based on the action matrix mapping, and sets Boolean logic conditions to ensure control security. Specifically, it includes the following steps: The scores and evidence strings of the six dimensions are summarized, and the weights of the confidence scores of the six dimensions are set according to the priority to calculate the comprehensive confidence score. The confidence interval is calculated and the evidence conflict is handled to generate the point confidence score and anomaly level. The exception level is mapped to a preset action, and a safety interlock is implemented based on Boolean logic conditions to output the processing action.

6. The six-dimensional intelligent operation and maintenance diagnostic system for oilfield gathering and transportation stations according to claim 5, characterized in that: The aggregated six-dimensional output scores and evidence strings are weighted according to priority allocation to calculate the comprehensive confidence score, calculate the confidence interval, handle evidence conflicts, and generate point confidence scores and anomaly levels. The specific steps are as follows: Weights are assigned to each dimension based on the reliability of the evidence. ,satisfy Calculate the overall confidence level ,in Score the confidence level; Calculate the mean and standard deviation of the six-dimensional rating vector S to set the confidence interval L, where: ; in, represent the mean, standard deviation, and sample size of the six-dimensional scoring vector S, respectively. This corresponds to a 95% confidence level. Output uncertainty index ; Through U pair Make corrections, including correcting the confidence level. ,when When the evidence is consistent, use it directly. As a basis for credibility, when This triggers conflict handling, including priority arbitration and uncertainty-driven degradation. Priority arbitration directly marks anomalies as abnormal when the design boundary dimension is determined to be anomaly. Uncertainty-driven degradation, on the other hand, involves... At that time, the credibility score anomaly level will be forcibly downgraded and labeled "high uncertainty". This is the uncertainty threshold; Calculate credibility score based on modified confidence level ,in These represent the extreme values ​​of the historical revised confidence level, when ≥0.8 and U< When the abnormal level is judged as normal, ≥0.8 and U≥ Or 0.5≤ When <0.8, the anomaly level is judged as suspicious. If the value is less than 0.5, the anomaly level is determined to be abnormal, and a conflict resolution record field is added to the evidence string.

7. The six-dimensional intelligent operation and maintenance diagnostic system for oilfield gathering and transportation stations according to claim 6, characterized in that: The process of mapping anomaly levels to preset actions, implementing safety interlocks based on Boolean logic conditions, and outputting processing actions specifically includes the following steps: For each determined anomaly level, an anomaly level-action mapping matrix is ​​constructed, showing the mapping relationships. , where K is a two-dimensional matrix, Level is the anomaly level, and Ac is the pre-defined action for the anomaly; At the same time, Boolean logic conditions are set to ensure that control actions are performed only under safe conditions, i.e. , These represent safety condition 1, safety condition 2, and hazardous condition, respectively. The data collection and execution receipts are used to calculate the achievement rate. When the achievement rate is below 0.9, an alternative strategy is triggered. These represent the numerical values ​​of the actual achieved result or state, and the preset target value, respectively.

8. A six-dimensional intelligent operation and maintenance diagnostic system for oilfield gathering and transportation stations according to claim 7, characterized in that, The feedback update module collects execution results and manually reviewed samples, and writes them back to the model library. It detects model performance changes based on KL divergence, updates the running boundary and residual threshold using incremental learning, and slowly deploys new parameters through shadow mode and A / B testing. It calculates model health based on false positive and false negative rates. The specific steps are as follows: Collect the execution results of the six-dimensional fusion processing module, manually reviewed samples, and new system operation data. Classify the data by type and write it back to the corresponding model library. The model library includes a success case library, a false positive sample library, a false negative sample library, and a basic historical database. Based on a fixed-period comparison of the model's distribution on the old and new data, where the old distribution P represents the historical data distribution on which the model was trained, and the new distribution Q represents the distribution of recently incoming online data, calculate the KL divergence: ; Where x is the value of the random variable, when > When a significant model drift is detected, model reconstruction is triggered. The divergence threshold; For operational boundaries and residual thresholds, an exponentially weighted moving average is used for online smooth updates. New parameters are slowly deployed using shadow mode and A / B testing. A model health score is calculated weekly. ,in The system identifies false alarms, missed alarms, and the total number of events. An alarm is triggered when H is below 0.

8.

9. A six-dimensional intelligent operation and maintenance diagnosis method for oilfield gathering and transportation stations, characterized in that, This method employs a six-dimensional intelligent operation and maintenance diagnostic system for oilfield gathering and transportation stations as described in any one of claims 1-8, and includes the following steps: S1. Access multi-source measurement data from various sensors in the oilfield gathering and transportation station, including flow rate, pressure, temperature, liquid level, water content, pump frequency / current, valve position / status position, furnace gas volume / damper, and station control alarm data. Clean, align, and extract features from the accessed measurement data to construct a computable single-point model. S2. Based on the obtained multi-source measurement data single-point model, the abnormal status and evidence are output for each point through a six-dimensional analysis link. The six dimensions include design boundary judgment, operation boundary judgment, process correlation verification, mechanism correlation verification, mechanism calculation and deduction, and big data reasoning analysis. Each link outputs a confidence score and evidence string, and the results of each link are summarized into a score vector. S3. Integrate the six-dimensional evidence output from the points, generate point credibility scores and anomaly levels, map the anomaly levels to preset actions based on the action matrix mapping, set Boolean logic conditions to ensure control safety, and output processing actions. S4. Collect execution results, manually review samples, and write them back to the model library. Detect model performance changes based on KL divergence, use incremental learning to update the running boundary and residual threshold, and slowly deploy new parameters through shadow mode and A / B testing. Calculate model health based on false positive rate and false negative rate.

Images