Oilfield station intelligent inspection method and device based on three-level architecture

Abstract

The application discloses a three-level architecture-based intelligent inspection method and device for an oilfield station, and relates to the technical field of oilfield gathering and transferring station inspection, which comprises the following steps: S1, the physical structure of the current oilfield gathering and transferring station is split into a region layer, a device layer and a point layer, an index table is established through a hash index to construct a three-level mapping table, the monitoring devices are multi-homedly bound, the sequence of the region layer, the device layer and the point layer is used to dequeue, a round of inspection sequence is constructed, a mechanism equation of a key process is constructed in a model layer and is linearized online to output a prediction residual error, the confidence is improved when the alarms are consistent, and when conflicts occur, a hierarchical Bayesian method is used to perform prior correction and threshold self-adaptation, and a physical event is output when the alarm is confirmed; and S2, multi-source alarms are generated based on sensor raw data. The application shortens the fault positioning path and improves the diagnosis accuracy by systematically managing the station objects, integrating multi-source alarms and adaptively determining faults.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of oilfield gathering and transportation station inspection, in particular to an oilfield station intelligent inspection method and device based on a three-level architecture. BACKGROUND

[0002] As the core hub of crude oil exploration and transportation, the oilfield gathering and transportation station covers a large number of physical objects such as tank farms, pump groups, separators and pipelines, and there is a complex process association between devices. The efficiency and accuracy of operation and inspection directly determine the safety and continuity of oilfield production.

[0003] However, in the traditional oilfield station inspection mode, the physical structure of the station lacks systematic hierarchical management, and the correlation between regions, devices and monitoring points is chaotic. When a fault alarm occurs, the maintenance personnel need to check one by one along the scattered information link, which not only takes a long time and is time-consuming, but also is easy to cause fault misjudgment or omission due to information omission, prolong the fault downtime, cause serious production loss, and the multi-source alarm data in the inspection process is fragmented, lacks unified modeling standard, the same fault may trigger multiple repeated or cross alarms, the maintenance personnel need to spend a lot of effort to filter effective information, it is difficult to quickly lock the fault source, especially for cross-device, cross-parameter mechanism fault, it is difficult to form a complete fault diagnosis link due to scattered data.

[0004] At present, there is no effective solution to the problems in the related art. SUMMARY

[0005] In view of the problems in the related art, the present application proposes an oilfield station intelligent inspection method and device based on a three-level architecture to overcome the above technical problems existing in the prior art.

[0006] To this end, the specific technical solutions adopted by the present application are as follows:

[0007] The oilfield station intelligent inspection method based on a three-level architecture comprises the following steps:

[0008] S1, the physical structure of the current oilfield gathering and transportation station is divided into region layer, device layer and point layer, an index table is established through hash index to build a three-level mapping table, and the monitoring devices are multi-attached and bound, and the region, device and point layer sequence is dequeued to build the current inspection sequence;

[0009] S2, generating multi-source alarms based on sensor raw data, including rule engine alarms, visual recognition alarms and mechanism model alarms, constructing an experience layer through IF-THEN rules, constructing mechanism equations of key processes in a model layer and outputting prediction residuals in linearization online, improving confidence when alarms are consistent, performing prior correction and threshold self-adaptation through hierarchical Bayesian when conflicts occur, and outputting physical events when alarms are confirmed;

[0010] S3, for the physical events that occur, visual recognition, empirical rules and mechanism model prediction are unified and packaged as five-tuples, and event deduplication and conflict resolution are realized through weight and priority combination, a unique fault object is output for the same physical event, the five-tuples include fault unique identifier, alarm source, abnormal description, severity level and timestamp, and a single alarm is output;

[0011] S4, when it is detected that the physical events exist across point position logical combinations, health state virtual points are generated, alarm evidences from rule engines, visual recognition and mechanism models are collected, an object mapping table is established to record the mapping relationship between health state virtual points and actual devices, and composite alarms are output.

[0012] As a preferred embodiment, the S1 includes the following steps:

[0013] S11, based on the design architecture of the current oilfield gathering and transferring station, the physical structure of the current oilfield gathering and transferring station is divided into a region layer, a device layer and a point layer, wherein the region layer is different functional regions of the current oilfield gathering and transferring station, the device layer is specific devices in different functional regions of the current oilfield gathering and transferring station, and the point layer is sensors and monitoring points associated with the devices;

[0014] S12, for the obtained region layer, device layer and point layer of the current oilfield gathering and transferring station, a three-level mapping is established through a hash index, and the monitoring devices are multi-ownership bound, and a round of inspection sequence is constructed according to the hierarchical order of region, device and point, including the following steps:

[0015] A unique identifier is allocated to each entity of the region layer, device layer and point layer, a hash function is used to take the identifier as input, and a modulo hash function is used to generate a corresponding hash value wherein identifier and hash table, respectively, and the identifier and its corresponding entity information are stored in the index table with the hash value as the index, wherein the entity information includes region name, device model and point type;

[0016] The index table is traversed and a three-level mapping table is established according to the hierarchical relationship of region, device and point, wherein the three-level mapping table is wherein represents the i-th region,​ represents the jth device in the ith region, represents the kth point of the jth device in the ith region;

[0017] For the monitoring device in the current oilfield gathering and transportation station, based on the monitoring range of the monitoring device, the region, device and point covered by each monitoring device are determined, the region, device and point covered by each monitoring device are recorded based on the three-level mapping table, and the elements in the three-level mapping table are dequeued according to the hierarchical order of region, device and point to construct the current round of inspection sequence.

[0018] As a preferred embodiment, S2 comprises the following steps:

[0019] S21, obtaining original data from each point of the current oilfield gathering and transportation station, and generating multi-source alarms from the original data, the multi-source alarms including rule engine alarms, visual recognition alarms and mechanism model alarms;

[0020] S22, adjusting the confidence according to the results of the rule engine alarms, the visual recognition alarms and the mechanism model alarms, performing prior correction and threshold self-adaptation with hierarchical Bayesian when the alarms are inconsistent, and confirming different alarms and outputting physical events when the confidence reaches the threshold.

[0021] As a preferred embodiment, S21 comprises the following steps:

[0022] S211, obtaining original data from each point of the current oilfield gathering and transportation station, and performing moving average filtering on the original data to remove noise, using IF-THEN rules to construct an experience layer in the rule engine alarm, and setting the rules as .

[0023] Meanwhile, based on the historical accuracy rate, the rule engine alarm confidence is dynamically calibrated for each rule , wherein , wherein represents the historical data weight coefficient, represents the historical accuracy rate, represents the static weight of the current rule, and when the rule engine alarm confidence is greater than the rule engine threshold, the rule engine alarm is output;

[0024] S212, obtaining image or video data of the current oilfield gathering and transportation station from the monitoring device, using a convolutional neural network to extract features from the image or video, comparing the extracted features with normal state features, and obtaining a current visual recognition alarm confidence , based on the extracted feature vector and the pre-stored normal state feature template By comparing and calculating, the distance to Mahalanobis was obtained. ,in The mean vector representing the set of normal feature samples. The covariance matrix representing the normal feature sample set is obtained by mapping using the Sigmoid function. ,in:

[0025] ;

[0026] in, Represents the decision boundary threshold. The slope parameter is based on image quality. right Make adjustments to obtain When the confidence level of the visual recognition alarm is greater than the visual recognition threshold, a visual recognition alarm is output.

[0027] S213. Construct mechanistic equations for key processes at the model level to describe the relationships between various parameters in the process. Use Taylor expansion to linearize the mechanistic equations online to obtain a linearized model. Compare the predicted parameter values ​​output by the linearized model with the actual measured values ​​to calculate the predicted residuals. Standardize the fluctuation of the original residuals r relative to the historical normal residuals to obtain statistics. ,in Represents the absolute value of the predicted residual. This represents the mean of recent historical residual data. The standard deviation of recent historical residual data is used to calculate the confidence level of the mechanistic model alarm based on statistics. ,in The cumulative distribution function represents the standard normal distribution. When the alarm confidence of the mechanism equation is greater than the threshold of the mechanism equation, a visual recognition alarm is output.

[0028] In a preferred embodiment, S22 includes the following steps:

[0029] S221. Based on the results of rule engine alarms, visual recognition alarms, and mechanism model alarms, when the three alarms output alarms simultaneously, appropriate fusion weights are assigned to each alarm source according to its real-time reliability and historical performance. ;

[0030] in These represent the real-time confidence level of the current alarm from that source and the historical weight base of that source, respectively. Then the overall confidence level for ,when The alarm is then acknowledged and a physical event is output, in which... representing the integrated confidence threshold;

[0031] S222, calculating the posterior probability according to the hierarchical Bayesian formula when the alarm results are inconsistent, adjusting the alarm threshold according to the updated model parameters and the posterior probability, specifically including the following steps:

[0032] calculating the posterior probability according to the hierarchical Bayesian formula , wherein represents the true state of the system, represents the alarm results from rules, vision, and models, respectively represent the prior probability, the likelihood function, and the posterior probability, and the posterior probability is calculated updating the cognition of the system state:

[0033] ;

[0034] , wherein represents the corrected result, which is based on updating the fusion weight , wherein is a similarity function, which returns 1 when the alarm is consistent with , and returns B when the alarm is inconsistent with , wherein the value of B ranges from 0 to 1, represents the new weight, and the updated integrated confidence is calculated according to repeating step S221 to obtain the updated integrated confidence, and confirming the alarm based on the updated integrated confidence threshold comparison result.

[0035] As a preferred embodiment, the S3 includes the following steps:

[0036] S31, collecting various physical events generated from visual recognition, experience rules, and mechanism model prediction, and constructing a five-tuple information for each physical event, including the following steps:

[0037] constructing a five-tuple information for each physical event, which includes a unique fault identifier, an alarm source, an abnormality description, a severity level, and a timestamp, and encapsulating the physical event as a five-tuple vector:

[0038] ;

[0039] S32, assigning corresponding weights to different alarm sources, comparing the information of each five-tuple to identify existing repeated physical events, and analyzing conflicting physical events based on the weights, and outputting a single alarm for a single fault.

[0040] As a preferred embodiment, the S32 includes the following steps:

[0041] S321, assign corresponding weights to different alarm sources, including visual recognition, empirical rules, mechanism model prediction, set priority for each physical event by comprehensively considering severity level and timestamp, and the specific steps are as follows:

[0042] ;

[0043] wherein, represents priority, respectively represent weight coefficients of severity level and timestamp, respectively represent severity level, timestamp and reference time of the current physical event;

[0044] S322, compare the information of each five-tuple, determine whether the event is repeated by comparing the abnormal description, compare the weight and priority for the repeated event, retain the event with high weight and priority and delete other repeated events;

[0045] S323, when there is a conflict between different five-tuple described abnormal situations, add the weight and priority of the two five-tuples, retain the five-tuple with high cumulative value and delete the five-tuple with low cumulative value, after event deduplication and conflict resolution, output the unique fault object in the form of single alarm for the same physical event.

[0046] As a preferred embodiment, the S4 comprises the following steps:

[0047] S41, continuously monitor the data of each point and the generated alarm information, including the physical quantity collected by the point and the alarm triggered by the rule engine and the abnormality found by visual recognition, define the cross-point logical combination rule to judge the cross-point abnormality;

[0048] S42, when there is a cross-point logical combination abnormality, a health state virtual point vector is generated to express the overall health state of the device under a specific cross-point logical combination, collect the alarm evidence from the rule engine, visual recognition and mechanism model and associate it with the current health state virtual point;

[0049] S43, create an object mapping table to record the mapping relationship between the health state virtual point and the actual device, add the generated health state virtual point to the object mapping table, establish a one-to-one correspondence between the virtual point and the corresponding device, and output the composite alarm.

[0050] The oilfield station intelligent inspection device based on the three-level architecture is applied to an oilfield station intelligent inspection method based on a three-level architecture, and includes a memory, a processor, and a program stored in the memory and capable of running on the processor.

[0051] The present application has the following advantages:

[0052] 1. The present application manages the entire object of the oilfield gathering and transportation station by a three-level mapping skeleton, so that the operation and maintenance inspection can be quickly positioned from top to bottom or from bottom to top to shorten the troubleshooting path. The physical structure of the oilfield gathering and transportation station is classified and split, and the hierarchical relationship of the region, equipment and point is expressed at the same time, so as to reduce the problem of long positioning path, easy to miss and easy to make mistakes when the alarm occurs, reduce the impact of failure on production, and improve the reliability and stability of the oilfield gathering and transportation station.

[0053] 2. The present application aggregates multiple source alarms by unified modeling of "five-tuple", generates a unique traceable "fault object", and introduces a "health state" virtual point to carry mechanism-based faults across points and parameters, so as to ensure that each alarm can be mapped to a specific equipment level and enter the life cycle management, so that the operation and maintenance personnel can quickly and accurately locate the fault source according to the "fault object", avoiding the process of repeatedly checking the fault in a complex system. The complex fault information is integrated and abstracted through the "health state" virtual point, and the complex fault can be accurately identified and diagnosed through monitoring and analysis of the virtual point.

[0054] 3. The present application introduces the hierarchical Bayesian inference framework by the interpretable rule and the online mechanism model residual, realizes the threshold self-adaptation and the confidence dynamic update, enhances the confidence when consistent and performs prior correction when contradictory, and has the interpretability and the robustness, so as to strengthen the judgment of the fault in the operation and maintenance process of the oilfield gathering and transportation station and enhance the functionality. BRIEF DESCRIPTION OF DRAWINGS

[0055] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the following will briefly introduce the drawings needed in the embodiments. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor.

[0056] Figure 1 is a flow chart of the oilfield station intelligent inspection method based on the three-level architecture according to the embodiment of the present application;

[0057] Figure 2 is a multi-alarm generation schematic diagram of the oilfield station intelligent inspection method based on the three-level architecture according to the embodiment of the present application. DETAILED DESCRIPTION

[0058] To further illustrate the embodiments, the present application provides accompanying drawings which form a part of the disclosure, which mainly serve to illustrate the embodiments, and can be used to explain the operating principles of the embodiments in conjunction with the relevant description. Those of ordinary skill in the art can understand other possible implementations and advantages of the present application by referring to these contents, and the components in the drawings are not drawn to scale, and similar component symbols are generally used to represent similar components.

[0059] According to an embodiment of the present application, a method and device for intelligent inspection of an oilfield station based on a three-level architecture are provided.

[0060] The present application will be further described in conjunction with the accompanying drawings and specific embodiments:

[0061] Embodiment 1:

[0062] As shown in the method for intelligent inspection of an oilfield station based on a three-level architecture according to an embodiment of the present application, the method comprises the following steps: Figure 1

[0063] S1, the physical structure of the current oilfield gathering and transportation station is split into a region layer, a device layer, and a point layer, a three-level mapping table is constructed by establishing an index table through a hash index, and the monitoring devices are multi-homed bound, and the region, device, and point layers are dequeued in order to construct a round of inspection sequence;

[0064] S11, based on the design architecture of the current oilfield gathering and transportation station, the physical structure of the current oilfield gathering and transportation station is divided into a region layer, a device layer, and a point layer, wherein the region layer is different functional regions of the current oilfield gathering and transportation station, the device layer is specific devices within different functional regions of the current oilfield gathering and transportation station, and the point layer is sensing and monitoring points associated with the devices;

[0065] It should be noted that the region layer can be different functional regions in the oilfield gathering and transportation station, such as oil storage areas and oil transportation areas, the device layer is specific devices in each region, such as pumps and valves, and the point layer is specific monitoring points on the devices, such as the positions of temperature sensors and pressure sensors. A three-level architecture of "region-device-point" is constructed to solve the problems of long positioning path and easy omission and error when an alarm occurs.

[0066] S12, for the obtained region layer, device layer, and point layer of the current oilfield gathering and transportation station, a three-level mapping is established through a hash index, and the monitoring devices are multi-homed bound, and the region, device, and point layers are dequeued in order to construct a round of inspection sequence, which specifically comprises the following steps:

[0067] ​A unique identifier is assigned to each entity of the area layer, the device layer and the point layer, a hash function is used to take the identifier as input, and the hash value is obtained by taking the modulus of the hash function A corresponding hash value is generated , wherein and represent the identifier and the hash table respectively, and the identifier and the corresponding entity information are stored in the index table with the hash value as the index, wherein the entity information includes the area name, the device model and the point type;

[0068] The index table is traversed and a three-level mapping table is established according to the hierarchical relationship of the area, the device and the point, wherein the three-level mapping table is , wherein represents the i-th area, represents the j-th device in the i-th area, represents the k-th point of the j-th device in the i-th area;

[0069] For the monitoring device in the current oilfield gathering and transportation station, based on the monitoring range of the monitoring device, the area, the device and the point covered by each monitoring device are determined, the area, the device and the point information covered by each monitoring device are recorded based on the three-level mapping table, and the elements in the three-level mapping table are dequeued in the order of the area, the device and the point, and a round of inspection sequence is constructed.

[0070] It should be noted that the three-level mapping table records the inclusion relationship between the area and the device, and the device and the point, and in the process of constructing the inspection sequence, each area is selected in turn from the area layer, and then the device in each area is selected in turn, and finally the point in each device is selected in turn, and the selected area, device and point are grouped into a round of inspection sequence in the order of dequeuing.

[0071] S2, multi-source alarm generation based on sensor raw data, including rule engine alarm, visual recognition alarm and mechanism model alarm, experience layer is constructed through IF-THEN rule, mechanism equation of key process is constructed in model layer and linearized output prediction residual is online, when alarm is consistent, confidence is improved, when contradiction occurs, prior correction and threshold self-adaptation are executed by hierarchical Bayesian, and physical event is output when alarm is confirmed;

[0072] S21, obtaining raw data from each point of the current oilfield gathering and transportation station, and generating multi-source alarms from the raw data, the multi-source alarms including rule engine alarms, visual recognition alarms and mechanism model alarms;

[0073] S211, obtaining raw data from each point of the current oilfield gathering and transportation station, and performing moving average filtering on the raw data to remove noise, and using IF-THEN rule to construct experience layer in rule engine alarm, and setting the rule as ;

[0074] Simultaneously, based on historical accuracy, dynamic calibration is performed to assign a rule engine alert confidence level to each rule. ,in ,in Represents the weighting coefficient of historical data. Represents historical accuracy. This represents the static weight of the current rule. When the rule engine alarm confidence score is greater than the rule engine threshold, a rule engine alarm is output.

[0075] It should be noted that the historical data weight coefficient is used to balance the proportion of historical accuracy and the static weight of the current rule in the final confidence calculation. The default value is 0.7. If the system has been running for a long time and the historical data is very reliable, the historical data weight coefficient can be flexibly adjusted. The historical accuracy is the accuracy of the current rule in triggering alarms in the past 30 days. The static weight is a fixed weight value pre-assigned by domain experts when defining the rule. Based on the expert's subjective judgment of the importance and reliability of the rule, the rule engine threshold is usually set to 0.6, but it can also be adjusted according to the actual situation.

[0076] S212. For the image or video data of the current oilfield gathering and transportation station obtained by the monitoring equipment, use a convolutional neural network to extract features from the image or video, compare the extracted features with the features of the normal state, and obtain the current visual recognition alarm confidence level. Based on the extracted feature vector With pre-stored normal state feature template By comparing and calculating, the distance to Mahalanobis was obtained. ,in The mean vector representing the set of normal feature samples. The covariance matrix representing the normal feature sample set is obtained by mapping using the Sigmoid function. ,in:

[0077] ;

[0078] in, Represents the decision boundary threshold. The slope parameter is based on image quality. right Make adjustments to obtain When the confidence level of the visual recognition alarm is greater than the visual recognition threshold, a visual recognition alarm is output.

[0079] It should be noted that, Represents the decision boundary threshold, when d= The confidence level is 0.5. The best value of the slope parameter needs to be debugged on actual data to determine the degree of change of the confidence from 0 to 1, in order to control the confidence When setting, an image data set containing only normal states needs to be collected, the normal sample set is processed by the trained convolutional neural network model, a feature vector is extracted for each normal picture, and the distance is calculated, and the is set as the mean + 3 times the standard deviation of the normal distance distribution, and the image quality score is to quantify the reliability of the input image, which is used for subsequent confidence correction, so as to reduce false positives caused by poor image quality. The image quality is composed of multiple dimensions of sub-scores, including sharpness, brightness and contrast. Sharpness is calculated by Laplacian variance method, brightness is calculated by calculating the mean value of all pixels of the gray image, and the mean value of brightness is mapped to the sub-score using the Gaussian function to obtain, and the contrast is calculated by calculating the standard deviation of the gray image. The larger the standard deviation, the more dispersed the pixel value distribution, and the higher the contrast. The standard deviation is mapped to the sub-score to obtain, and the image quality is calculated according to the weights 0.3, 0.3 and 0.4. The visual recognition threshold is usually set to 0.6, which can also be adjusted according to actual conditions.

[0080] S213, the mechanism equation of the key process is constructed in the model layer to describe the relationship between various parameters in the process. The Taylor expansion is used to linearize the mechanism equation online to obtain a linear model. The predicted parameter value output by the linear model is compared with the actual measured value, and the prediction residual is calculated. The fluctuation of the original residual r relative to the historical normal residual is standardized to obtain a statistical quantity , wherein represents the absolute value of the prediction residual, represents the mean of the recent historical residual data, represents the standard deviation of the recent historical residual data, and the mechanism model alarm confidence is calculated based on the statistical quantity , wherein represents the cumulative distribution function of the standard normal distribution, and when the mechanism equation alarm confidence is greater than the mechanism equation threshold, the visual recognition alarm is output.

[0081] It should be noted that when the mechanism model confidence is actually calculated, and need to change over time, and a residual data window in the last N sampling periods is maintained by the system , and , The judgment benchmark of the mechanism model confidence is dynamically adjusted according to the recent running state of the system, and the mechanism equation threshold is usually set to 0.6, which can also be adjusted according to the actual situation.

[0082] S22, according to the results of the rule engine alarm, visual recognition alarm and mechanism model alarm, the confidence is adjusted, when the alarms are consistent, the confidence is adjusted, when the alarms are inconsistent, the prior correction and threshold self-adaptation are performed according to the hierarchical Bayesian method, when the confidence reaches the threshold, the different alarms are confirmed and the physical events are outputted;

[0083] S221, according to the results of the rule engine alarm, visual recognition alarm and mechanism model alarm, when the three alarms are outputted synchronously, the appropriate fusion weight is assigned to each alarm source according to the real-time reliability and historical performance of the alarm source ;

[0084] Among them respectively represent the real-time confidence of this alarm of the source and the historical weight base of the source, , the comprehensive confidence is , when , the alarm is confirmed and the physical event is outputted, wherein represents the comprehensive confidence threshold;

[0085] It should be noted that represents the comprehensive confidence threshold, and the value is usually set to 0.75, when the confidence of the three sources is high and consistent, the fusion result will be a value higher than the average value of any single source, so as to achieve the confidence improvement effect.

[0086] S222, when the alarm results are inconsistent, the posterior probability is calculated according to the hierarchical Bayesian formula, the alarm threshold is adjusted according to the updated model parameters and the posterior probability, and the specific steps include the following steps:

[0087] The posterior probability is calculated according to the hierarchical Bayesian formula , wherein represents the true state of the system, represents the alarm results from rules, vision and model, respectively represent the prior probability, the likelihood function and the posterior probability, and the posterior probability updates the cognition of the system state:

[0088] ;

[0089] Among them, represents the corrected result, which is based on Updating the fusion weight , wherein is a similarity function, returning 1 when the alarm is consistent with , and returning B when the alarm is inconsistent with , wherein B ranges from 0 to 1, represents a new weight, and repeating step S221 to obtain an updated comprehensive confidence, and confirming the alarm based on a comparison result of the updated comprehensive confidence threshold.

[0090] It should be noted that when the alarm is inconsistent with , B is returned, wherein B ranges from 0 to 1, and is usually set to 0.3, meaning that the source is not very reliable this time, and its weight will be multiplied by 0.3, wherein , represents a new weight, which will be used for the next fusion calculation after adjustment, and the weight of a good source is retained, and the weight of a poor source is reduced, and the weight is dynamically adjusted according to the consistency of each source and the Bayesian inference result after a conflict occurs, so as to realize the updating confirmation of the alarm, represents the true state of the system, which needs to be set in advance according to specific conditions, for example = 0 represents “device health”, = 1 represents “pump leakage”, = 2 represents “motor overheating”, and the like, and different alarms are output for different physical events. Embodiment 2

[0091] S3, for the physical event that occurs, unified packaging of visual recognition, experience rule and mechanism model prediction into a five-tuple, and realizing event deduplication and conflict resolution through weight and priority combination, outputting a unique fault object for the same physical event, the five-tuple including fault unique identification, alarm source, abnormal description, severity level and timestamp, and outputting a single alarm;

[0092] S31, collecting various physical events generated from visual recognition, experience rule and mechanism model prediction, and constructing five-tuple information for each physical event, including the following steps:

[0093] constructing five-tuple information for each physical event, including fault unique identification, alarm source, abnormal description, severity level and timestamp, and packaging the physical event into a five-tuple vector:

[0094]

[0095] ;​​

[0096] It should be noted that, respectively represent the unique identifier of the fault, the alarm source, the abnormal description, the severity level and the timestamp, wherein the unique identifier of the fault is assigned to each physical event for subsequent distinction and management, the alarm source indicates that the physical event is generated by point abnormality, visual recognition, experience rule or mechanism model prediction, the abnormal description describes the specific abnormal situation of the physical event, for example, "temperature sensor reading exceeds the normal range", "obvious damage appears on the appearance of the equipment", etc., the severity level needs to be divided into severity levels 1, 2 and 3 according to the influence degree and emergency degree of the event by consulting experts in the relevant field combined with the current oil field gathering and transportation involved image, and the timestamp records the specific time when the physical event occurs.

[0097] S32, assign corresponding weights to different alarm sources, compare the information of each five tuple to identify the repeated physical events, and analyze the conflict physical events based on the weights, and output a single alarm for a single fault;

[0098] S321, assign corresponding weights to different alarm sources, including visual recognition, experience rule and mechanism model prediction, set the priority of each physical event by comprehensively considering the severity level and timestamp, and the specific steps are as follows:

[0099] ;

[0100] wherein, represents the priority, respectively represent the weight coefficients of the severity level and the timestamp, respectively represent the severity level, the timestamp and the reference time of the current physical event;

[0101] It should be noted that, assigning corresponding weights to different alarm sources needs to be based on historical data, expert experience or the reliability of each alarm source, wherein the mechanism model prediction has higher reliability based on physical principles, so it is assigned a higher weight, the visual recognition is greatly affected by environmental factors, so it is assigned a lower weight, and the experience rule is assigned a medium weight, the sum of the weights of the three is 1, and the priority of the event with high severity level and recent occurrence time is set to be high based on the priority calculation formula.

[0102] S322, compare the information of each five tuple, determine whether the event is repeated by comparing the abnormal description, compare the weight and priority for the repeated event, keep the event with high weight and priority and delete other repeated events;

[0103] It should be noted that if the anomaly descriptions meet a certain similarity threshold, the two events are considered to be duplicates. The similarity threshold needs to be set according to the actual equipment involved in the current alarm at the oilfield gathering and transportation station. The cosine similarity is calculated and compared with the similarity threshold to determine whether the events are duplicates.

[0104] S323. When there are conflicts in the abnormal situations described by different quintuples, the weights and priorities of the two quintuples are accumulated, the quintuple with the higher accumulated value is retained and the quintuple with the lower accumulated value is deleted. After event deduplication and conflict resolution, for the same physical event, a unique fault object is output in the form of a single alarm.

[0105] S4. When a physical event is detected to have a cross-point logical combination, a virtual health point is generated, alarm evidence from the rule engine, visual recognition and mechanism model is collected, an object mapping table is established to record the mapping relationship between the virtual health point and the actual device, and a composite alarm is output at the same time.

[0106] S41. Continuously monitor the data and alarm information generated at each point, including the physical quantities collected at the point, alarms triggered by the rule engine, and anomalies detected by visual recognition. Define cross-point logical combination rules to determine cross-point anomalies.

[0107] It should be noted that the defined cross-point logical combination rules are used to determine cross-point anomalies. Specifically, this includes assuming there are q points, each with corresponding monitoring data. The cross-point logical combination rule is represented as a logical expression. For example, the rule is "data at point 1". Greater than the threshold And the data at point 2 If the threshold is less than 1, then the logical expression is: ,when If the result is true, it indicates that there is an anomaly in the cross-point logic combination.

[0108] S42. When there is an anomaly in the cross-point logical combination, a health status virtual point vector is generated to express the overall health status of the device under a specific cross-point logical combination. Alarm evidence from the rule engine, visual recognition and mechanism model is collected and associated with the current health status virtual point.

[0109] It should be noted that alarm evidence collected from the rule engine, visual recognition, and mechanism model is associated with the current health status virtual point. This evidence is related to the current cross-point logical combination anomaly, such as specific data values ​​provided by sensors, specific rule content triggered by the rule engine, and images or feature information recognized by vision. Associating it with the health status virtual point makes it convenient to view and trace the source later.

[0110] S43, create an object mapping table for recording the mapping relationship between the health state virtual point and the actual device, add the generated health state virtual point to the object mapping table, and establish a one-to-one correspondence between the virtual point and the corresponding device, and output the composite alarm.

[0111] It should be noted that the object mapping table is U, and the device corresponding to the virtual point I is L, so that the virtual point is added to the object mapping table, which can be expressed as U(I)=L, that is, the mapping table U records the correspondence between the virtual point I and the device L, which can facilitate subsequent searching of related virtual point information according to the device, or locating to a specific device according to the virtual point information. The composite alarm information needs to be combined with the characteristics of the cross-point logic to output accurate fault information, including the involved devices, sensors, etc.

[0112] Embodiment 3:

[0113] The oilfield station intelligent inspection device based on the three-level architecture is applied to an oilfield station intelligent inspection method based on a three-level architecture, and includes a memory, a processor, and a program stored in the memory and executable on the processor. When the processor executes the program, the steps in the oilfield station intelligent inspection method based on the three-level architecture are implemented.

[0114] In summary, the present application manages the entire object of the oilfield gathering and transportation station by using a three-level mapping skeleton, so that the operation and maintenance inspection can be quickly positioned from top to bottom or from bottom to top to shorten the troubleshooting path. By classifying the physical structure of the oilfield gathering and transportation station, the hierarchical relationship between the region, the device and the point is expressed, the problem of long positioning path, easy to miss and easy to make mistakes when the alarm occurs is reduced, the influence of the fault on the production is reduced, the reliability and stability of the oilfield gathering and transportation station are improved, the "five-tuple" unified modeling is used to aggregate multi-source alarms, the unique traceable "fault object" is generated, and the "health state" virtual point is introduced to carry the mechanism fault of cross-point and cross-parameter. Ensure that each alarm can be mapped to a specific device level and enter the life cycle management, so that the operation and maintenance personnel can quickly and accurately locate the fault source according to the "fault object", avoiding the process of repeatedly checking the fault in a complex system. The complex fault information is integrated and abstracted through the "health state" virtual point, the complex fault can be accurately identified and diagnosed through the monitoring and analysis of the virtual point, the threshold self-adaptation and confidence dynamic update are realized by introducing the explainable rule and online mechanism model residual into the hierarchical Bayesian inference framework, the confidence is improved when consistent, and the prior correction is performed when contradictory, combining the explainability and robustness, and strengthening the judgment of the fault in the operation and maintenance process of the oilfield gathering and transportation station.

[0115] The above merely provides the preferred embodiment of the present application, and is not used to limit the present application. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application shall be included in the protection scope of the present application.

Claims

1. A method for intelligent inspection of an oilfield station based on a three-level architecture, characterized in that, The method includes the following steps: S1. The physical structure of the current oilfield gathering and transportation station is divided into a regional layer, an equipment layer, and a location layer. An index table is built using a hash index to construct a three-level mapping table. At the same time, the monitoring equipment is bound to multiple affiliations and is queued according to the order of the regional, equipment, and location layers to construct the inspection sequence for this round. S2. Generate multi-source alarms based on raw sensor data, including rule engine alarms, visual recognition alarms and mechanism model alarms. Construct an experience layer through IF-THEN rules, construct the mechanism equations of key processes in the model layer and linearize the prediction residuals online. Increase confidence when alarms are consistent, perform prior correction and threshold adaptation with hierarchical Bayes when contradictions occur, and output physical events when alarms are confirmed. S3. For physical events, visual recognition, empirical rules and mechanism model prediction are uniformly encapsulated into a five-tuple, and event deduplication and conflict resolution are achieved by merging weights and priorities. A unique fault object is output for the same physical event. The five-tuple includes a unique fault identifier, alarm source, anomaly description, severity level and timestamp, and a single alarm is output. S4. When a physical event is detected to have a cross-point logical combination, a virtual health point is generated, alarm evidence from the rule engine, visual recognition and mechanism model is collected, an object mapping table is established to record the mapping relationship between the virtual health point and the actual device, and a composite alarm is output.

2. The intelligent inspection method for oilfield station based on three-level architecture according to claim 1, characterized in that, S1 includes the following steps: S11. Based on the current design architecture of the oilfield gathering and transportation station, the physical structure of the current oilfield gathering and transportation station is divided into a regional layer, an equipment layer, and a point layer. The regional layer consists of different functional areas of the current oilfield gathering and transportation station, the equipment layer consists of specific equipment within different functional areas of the current oilfield gathering and transportation station, and the point layer consists of sensing and monitoring points associated with the equipment. S12. For the obtained regional, equipment, and location layers of the current oilfield gathering and transportation station, a three-level mapping is established using a hash index, and multiple affiliations are bound to the monitoring equipment. The equipment is then dequeued according to the order of the regional, equipment, and location layers to construct the inspection sequence for this round. The specific steps include: A unique identifier is assigned to each entity of the area layer, the device layer and the point layer, and the identifier is taken as input of a hash function by taking modulo of the hash function A corresponding hash value is generated Wherein The identifier and the hash table are represented respectively, and the identifier and the corresponding entity information are stored in the index table with the hash value as index, wherein the entity information includes area name, device model and point type. The index table is traversed and a three-level mapping table is established according to the hierarchical relationship of the area, the device and the point, wherein the three-level mapping table is wherein represents the i-th area, represents the j-th device in the i-th area, represents the k-th point of the j-th device in the i-th area. For the monitoring equipment in the current oilfield gathering and transportation stations, based on the monitoring range of the monitoring equipment, the area, equipment and location covered by each monitoring equipment are determined. On the basis of the three-level mapping table, the area, equipment and location information covered by each monitoring equipment is recorded. According to the hierarchical order of area, equipment and location, the elements in the three-level mapping table are dequeued to construct the inspection sequence for this round.

3. The intelligent inspection method for oilfield station based on three-level architecture according to claim 1, characterized in that, S2 includes the following steps: S21. Obtain raw data from various points in the current oilfield gathering and transportation station, and generate multi-source alarms from the raw data. Multi-source alarms include rule engine alarms, visual recognition alarms, and mechanism model alarms. S22. Based on the results of rule engine alarms, visual recognition alarms, and mechanism model alarms, adjust the confidence level. When the alarms are consistent, adjust the confidence level. When the alarms are inconsistent, perform prior correction and threshold adaptation using hierarchical Bayes. When the confidence level reaches the threshold, confirm the different alarms and output the physical event.

4. The intelligent inspection method for oilfield station based on three-level architecture according to claim 3, characterized in that, S21 includes the following steps: S211, obtain the original data for each point of the current oilfield gathering and transferring station, perform moving average filtering on the original data to remove noise, use IF-THEN rules to construct an experience layer in the rule engine alarm, set the rules as ; Meanwhile, based on historical accuracy rate dynamic calibration, each rule is given a rule engine alarm confidence wherein wherein represents a historical data weight coefficient, represents a historical accuracy rate, represents a static weight of the current rule, and when the rule engine alarm confidence is greater than a rule engine threshold value, a rule engine alarm is output. S212, for the current oilfield gathering and transferring station image or video data obtained by the monitoring device, using a convolutional neural network to extract features from the image or video, comparing the extracted features with the normal state features, obtaining the current visual recognition alarm confidence , based on the extracted feature vector Compare with the pre-stored normal state feature template , calculate the Mahalanobis distance , wherein The mean vector of the normal feature sample set represents The covariance matrix of the normal feature sample set is mapped using the Sigmoid function to obtain , wherein: ； wherein, representing a decision boundary threshold, is a slope parameter, while based on image quality to adjustment, obtain When the visual recognition alarm confidence is greater than the visual recognition threshold, output the visual recognition alarm. S213, constructing a mechanism equation of the key process at the model layer to describe the relationship between various parameters in the process, using Taylor expansion to perform online linearization on the mechanism equation to obtain a linearization model, comparing the predicted parameter value output by the linearization model with the actual measured value, calculating a prediction residual, and standardizing the fluctuation of the original residual r relative to the historical normal residual to obtain a statistic wherein represents an absolute value of the prediction residual, represents a mean value of the recent historical residual data, represents a standard deviation of the recent historical residual data, and a mechanism model alarm confidence is calculated based on the statistic wherein represents a cumulative distribution function of a standard normal distribution, and a visual recognition alarm is output when the mechanism equation alarm confidence is greater than a mechanism equation threshold value.

5. The intelligent inspection method for oilfield station based on three-level architecture according to claim 4, characterized in that, The S22 comprises the following steps: S221、According to the results of the rule engine alarm, visual recognition alarm and mechanism model alarm, when the three are synchronized to output alarms, appropriate fusion weights are assigned to each alarm source according to its real-time reliability and historical performance ; wherein respectively represent the real-time confidence of the source for this alert and the historical weight base of the source, then the integrated confidence is when then the alert is confirmed and a physical event is output, wherein represents an integrated confidence threshold; S222, when the alarm results are inconsistent, calculating the posterior probability according to the hierarchical Bayesian formula, adjusting the alarm threshold according to the updated model parameters and the posterior probability, and specifically comprising the following steps: The posterior probability is calculated according to the hierarchical Bayesian formula wherein represents the true state of the system, represents the alarm result from rules, vision, model, respectively represent the prior probability, the likelihood function and the posterior probability, and the cognition of the system state is updated through the calculated posterior probability ​ ； wherein, representing the corrected result, based on updating the fusion weight wherein, is a similarity function, returning 1 when the alarm is consistent with and B, wherein B has a value ranging from 0 to 1, when the alarm is inconsistent with , representing the new weight, according to repeating step S221 to obtain an updated comprehensive confidence, and confirming the alarm based on a comparison result of the updated comprehensive confidence threshold.

6. The intelligent inspection method for oilfield station based on three-level architecture according to claim 5, characterized in that, The S3 comprises the following steps: S31, collecting various physical events generated from visual recognition, experience rules and mechanism model prediction, and constructing a five-tuple information for each physical event, comprising the following steps: Constructing a five-tuple information for each physical event, which includes a unique fault identifier, an alarm source, an abnormality description, a severity level and a timestamp, and encapsulating the physical event as a five-tuple vector: ； S32, assigning corresponding weights to different alarm sources, comparing the information of each five-tuple to identify repeated physical events, and analyzing the conflict physical events based on the weights, and outputting a single alarm for a single fault.

7. The intelligent inspection method of oilfield station based on three-level architecture according to claim 6, characterized in that, The S32 comprises the following steps: S321, assigning corresponding weights to different alarm sources, including visual recognition, experience rules and mechanism model prediction, and setting a priority for each physical event by comprehensively considering the severity level and the timestamp, and specifically comprising the following steps: ； wherein, represents a priority, respectively represent a weight coefficient of a severity level and a timestamp, respectively represent a severity level, a timestamp and a reference time of a current physical event; S322, comparing the information of each five-tuple, determining whether the events are repeated by comparing the abnormality descriptions, comparing the weights and priorities for the repeated events, and retaining the events with high weights and priorities and deleting other repeated events; S323, when there are conflicts between different five-tuple described abnormal situations, adding the weights and priorities of the two five-tuples, retaining the five-tuple with a high cumulative value and deleting the five-tuple with a low cumulative value, and after event deduplication and conflict analysis, outputting a unique fault object in the form of a single alarm for the same physical event.

8. The intelligent inspection method of oilfield station based on three-level architecture according to claim 6, characterized in that, The S4 comprises the following steps: S41, continuously monitoring the data of each point and the generated alarm information, including the physical quantities collected by the point and the alarms triggered by the rule engine and the abnormalities found by visual recognition, and defining a cross-point logical combination rule to determine the cross-point abnormality; S42, when there is a cross-point logical combination abnormality, a health state virtual point vector is generated to express the overall health state of the device under a specific cross-point logical combination, and alarm evidence from the rule engine, visual recognition and mechanism model is collected and associated with the current health state virtual point; S43, creating an object mapping table to record the mapping relationship between the health state virtual point and the actual device, adding the generated health state virtual point to the object mapping table, and establishing a one-to-one correspondence between the virtual point and the corresponding device, and outputting a composite alarm.

9. The intelligent inspection device for oilfield station based on three-level architecture, characterized in that, The device is applied to an oilfield station intelligent inspection method based on a three-level architecture, comprising a memory and a processor: The memory is used for non-transitory storage of computer readable instructions; The processor is used for running the computer readable instructions; Wherein, the computer readable instructions are run by the processor, the method of any one of claims 1-8 is executed.

Images