Oil and petrochemical whole factory area weak electric integrated intelligent operation and maintenance management method and system based on edge computing

Abstract

The application discloses an oil and petrochemical plant area weak current integrated intelligent operation and maintenance management method and system based on edge computing, and relates to the weak current integrated operation and maintenance technical field of an oil and petrochemical plant area. The method is executed by an edge computing node deployed in a plant area, abnormal signals generated in real time by multiple weak electronic systems are acquired, the abnormal signals are subjected to space-time correlation to generate candidate cause-effect pairs, the normal occurrence probability of each type of abnormal event is maintained, the cause-effect promotion degrees of the two directions of the candidate cause-effect pairs are calculated, the cause-effect relationship direction is determined, a cause-effect relationship edge is obtained, the cause-effect relationship edges are concatenated into a cause-effect event chain, and the cause-effect event chain is output after being pruned through physical logic constraints. The application realizes automatic reasoning from multiple weak electronic system abnormal signals to a cause-effect event chain.

Description

Technical Field

[0001] This invention relates to the field of low-voltage integrated operation and maintenance technology in oil and petrochemical plants. More specifically, this invention relates to a method and system for intelligent operation and maintenance management of low-voltage integrated systems in oil and petrochemical plants based on edge computing. Background Technology

[0002] Oil and petrochemical plant areas are equipped with various low-voltage electronic systems, including video surveillance, vibration fiber optic perimeter systems, access control systems, fire alarm systems, and environmental monitoring systems. Each subsystem operates independently and generates alarm information separately. Currently, the operation and maintenance management of these low-voltage electronic systems mainly adopts a data aggregation approach, centralizing the alarm data from each subsystem to a unified platform for display. China Telecom Corporation Limited's authorized patent CN115150250B discloses a method and device for locating the root cause of IoT anomalies based on causal learning. This method locates the root cause by performing causal learning on abnormal alarm data in the IoT system, representing a post-event analysis solution for alarms that have already occurred.

[0003] While the above methods achieve centralized alarm data, alarms from different subsystems remain semantically isolated. The system can inform maintenance personnel that a certain alarm occurred at a certain location and time, but it cannot automatically reveal potential causal relationships between alarms from different subsystems. When a real anomaly occurs in the plant, alarms are often triggered sequentially in multiple low-voltage electronic systems. Faced with a large amount of fragmented alarm information, maintenance personnel need to rely on personal experience to manually analyze the correlations between alarms, determine which alarms are different manifestations of the same event, and identify the causal logic between alarms. This process has limited efficiency and is prone to missing key causal clues when alarms are dense. Therefore, to address the above issues, a method and system for integrated intelligent operation and maintenance management of low-voltage electrical systems across the entire oil and petrochemical plant based on edge computing is proposed. Summary of the Invention

[0004] To overcome the aforementioned deficiencies of the prior art, embodiments of the present invention provide a method and system for intelligent operation and maintenance management of low-voltage electrical systems across the entire oil and petrochemical plant based on edge computing. This system aims to address the problems of isolated alarm semantics in various subsystems and the inability to automatically discover cross-system causal relationships in existing low-voltage electrical integrated operation and maintenance methods, as well as the problem of limited processing efficiency caused by relying on manual experience to analyze fragmented alarms.

[0005] To achieve the above objectives, the present invention provides the following technical solution: A method for integrated intelligent operation and maintenance management of low-voltage electrical systems across an entire oil and petrochemical plant, based on edge computing and executed by edge computing nodes deployed throughout the plant, includes the following steps: S1. Acquire abnormal signals generated in real time by multiple weak electronic systems within the oil and petrochemical plant area. The weak electronic systems include video surveillance system, vibration fiber optic perimeter system, access control system, fire alarm system and environmental monitoring system. Each abnormal signal includes the type of its subsystem, abnormal feature value, timestamp and logical spatial coordinates, so that abnormal events from different weak electronic systems can be processed in a unified spatiotemporal framework.

[0006] S2. For any weak electrical system signal identified as an anomalous change at the current moment, take it as a seed signal. Determine the basic scale of the window according to the subsystem type to which the seed signal belongs, and determine the window scaling factor according to the anomalous amplitude of the seed signal. The time span and spatial radius of the spatiotemporal correlation window are obtained by multiplying the basic scale of the window and the window scaling factor. Anomalous signals from other weak electrical systems whose timestamps and logical spatial coordinates fall within the spatiotemporal correlation window are combined with the seed signal to form candidate causal pairs.

[0007] The scale of the aforementioned spatiotemporal correlation window is adaptively adjusted according to the type and severity of the seed signal in order to capture possible cross-system causal correlation signals.

[0008] S3. Maintain the normal occurrence probability for each type of weak current system abnormal event. This normal occurrence probability is based on the independent occurrence record of the abnormal event when it does not form a candidate causal pair with other weak current system abnormal signals, and is dynamically updated by incremental Bayesian estimation with forgetting.

[0009] Among them, the weight of historical observation data decays over time, so that the probability of normal occurrence follows the changes in the plant's operating status.

[0010] S4. For each candidate causal pair, calculate the causal lift in both directions. By comparing the causal lift in both directions, determine the direction of the causal relationship and obtain the causal relationship edge with direction label and confidence weight. The causal lift is the ratio of the conditional probability of an event after another event occurs to the normal probability of the event.

[0011] When the causal lift in one direction is significantly higher than that in the opposite direction, it is determined that there is a causal relationship in that direction, and the value of the lift is used as the confidence level of the causal relationship.

[0012] S5. Connect multiple causal relationship edges in chronological order and causal direction to form a causal event chain. After pruning by preset physical logic constraints, output a structured causal event chain object. The causal event chain object includes root cause event identifiers, event node sequences, and confidence weights for each causal relationship edge.

[0013] The physical logic constraints are used to eliminate statistically significant but physically invalid causal relationships.

[0014] The above method transforms the problem of causal discovery across low-voltage systems into probability boosting calculation, without the need for predefined causal rules or labeled samples. It completes the reasoning process from multi-source abnormal signals to causal event chains in real time through edge nodes.

[0015] Furthermore, in step S2, the basic window scale corresponding to the diffusion-type anomaly event is larger than the basic window scale corresponding to the fixed-point anomaly event, so as to match the requirements of different physical propagation characteristics for the spatiotemporal correlation range.

[0016] The abnormal amplitude is positively correlated with the window scaling factor, and the higher the severity of the abnormality, the larger the spatiotemporal correlation search range.

[0017] Furthermore, step S2 also includes: monitoring the number of candidate causal pairs generated by a single seed signal.

[0018] When the number of candidate causal pairs exceeds a set upper limit, the window scaling factor corresponding to this type of abnormal event is reduced, thereby suppressing noise introduced by too many irrelevant signals.

[0019] When the number of candidate causal pairs is 0, increase the window scaling factor corresponding to this type of abnormal event to avoid missing potential causal relationships due to an excessively small window.

[0020] Furthermore, step S3 includes: maintaining three parallel, independent occurrence recording sliding windows—short-term window, medium-term window, and long-term window—for each type of weak current system abnormal event. The windows with different time granularities capture recent changes, daily cycle patterns, and long-term trends, respectively.

[0021] For each independent occurrence record within a sliding window, the incremental Bayesian estimation with forgetting is used to obtain the probability estimate for that window, wherein each update first adjusts the prior distribution hyperparameters of the Bayesian estimation. and Apply exponential decay, with a decay factor of 1. and Then, the new observation count within the current window is added on top of it. The smaller the decay factor, the faster the weight of the historical data decays.

[0022] The probability estimates corresponding to the three sliding windows are weighted and fused together. The weights are adaptively adjusted according to the number of observed samples in each window. The window with a larger sample size receives a higher weight, thus obtaining the probability of normal occurrence.

[0023] Further, step S4 includes: for the first event X and the second event Y in the candidate causal pair, definition , , in Let Y be the conditional probability of X after X occurs. Let Y be the probability of the normal occurrence. Let X be the conditional probability of Y after Y occurs. Let X be the normal probability of occurrence. Causal lift measures the degree to which the occurrence of one event increases the probability of the occurrence of another event.

[0024] calculate Directional asymmetry ratio Used to compare the degree of asymmetry in causal effects in two directions.

[0025] when Greater than the preset upper threshold and When the confidence threshold is greater than 1, it indicates that the lift in the X→Y direction is significantly higher than that in the reverse direction and the absolute strength is sufficient. Therefore, the causal relationship is determined to be X pointing to Y. As a confidence weight.

[0026] when Less than the preset lower threshold and When the confidence threshold is greater than 1, the causal relationship is determined to be Y pointing to X, and then... As a confidence weight.

[0027] when When the value falls between the preset lower threshold and the preset upper threshold, the lift in both directions is close, indicating a symmetrical coupling relationship.

[0028] Furthermore, step S5 includes: based on the determined causal relationship edges, searching forward along the time axis for subsequent relationship edges with the current result event as the cause for forward growth, and searching backward along the time axis for preceding relationship edges with the current cause event as the result for reverse tracing, alternating until both ends of the link can no longer be extended, to obtain the original causal event chain. This bidirectional growth method connects the scattered causal relationship edges into a complete event evolution link.

[0029] The system queries a pre-defined physical constraint knowledge base, which stores the causal role attributes and causal pairing taboos of each weak electronic system. The causal role attributes represent that a subsystem event can only be a cause, only an effect, or both. The causal pairing taboos are subsystem event association pairs that do not conform to physical causal logic.

[0030] When a link segment in the original causal event chain violates the causal role attribute or hits the causal pairing taboo, it is pruned. The pruned causal event chain is output as the structured causal event chain object. The physical logic constraints are used to filter and statistically analyze causal relationships that are related but not physically valid.

[0031] Furthermore, the method also includes receiving confirmation feedback, negative feedback, or correction feedback from operations and maintenance personnel regarding the output causal event chain object.

[0032] The feedback results are converted into virtual counts and injected incrementally into the Bayesian estimators corresponding to the normal occurrence probability and the conditional probability, so that the probability estimation of the system is optimized as operation and maintenance experience is accumulated.

[0033] The physical constraint knowledge base is updated based on the causal event chain pattern that has been repeatedly denied, so that the system avoids repeatedly outputting causal chains that have been confirmed to be wrong.

[0034] An intelligent operation and maintenance management system for low-voltage electrical systems across an entire oil and petrochemical plant, based on edge computing and deployed within edge computing nodes within the plant, includes: The candidate causal pair generation module acquires real-time abnormal signals from multiple weak electronic systems within the oil and petrochemical plant area. These weak electronic systems include video surveillance systems, vibration fiber optic perimeter systems, access control systems, fire alarm systems, and environmental monitoring systems. Any weak electronic system signal identified as an abnormal mutation is used as a seed signal. The basic window scale is determined based on the subsystem type to which the seed signal belongs, and the window scaling factor is determined based on the abnormal amplitude of the seed signal. The time span and spatial radius of the spatiotemporal correlation window are obtained by multiplying the basic window scale and the window scaling factor. Other subsystem abnormal signals whose timestamps and logical spatial coordinates fall within this spatiotemporal correlation window are combined with the seed signal to form candidate causal pairs.

[0035] The scale of the aforementioned spatiotemporal correlation window adaptively adjusts according to the type and severity of the seed signal.

[0036] The normal probability maintenance module maintains the normal occurrence probability for each type of abnormal event in the low-voltage system. This normal occurrence probability is based on the independent occurrence records of the abnormal event when it does not form a candidate causal pair. It is dynamically updated through incremental Bayesian estimation with forgetting, where the weight of historical observation data decays over time.

[0037] The causal direction determination module calculates the causal lift of two directions for each candidate causal pair, determines the causal relationship direction by comparing the causal lift of the two directions, and outputs the causal relationship edge with direction label and confidence weight. The causal lift is the ratio of the conditional probability of an event after another event occurs to the normal probability of the event.

[0038] The causal chain generation module connects multiple causal relationship edges in chronological order and causal direction to form a causal event chain. After pruning by preset physical logic constraints, it outputs a structured causal event chain object, which includes root cause event identifiers, event node sequences, and confidence weights for each causal relationship edge.

[0039] The physical logic constraints are used to eliminate statistically significant but physically invalid causal relationships.

[0040] Furthermore, the normal probability maintenance module maintains three parallel, independent occurrence recording sliding windows—short-term window, medium-term window, and long-term window—for each type of abnormal event in the low-voltage system. The windows with different time granularities capture recent changes, daily cycle patterns, and long-term trends, respectively.

[0041] For each independent occurrence record within a sliding window, the incremental Bayesian estimation with forgetting is used to obtain the probability estimate for that window, wherein each update first adjusts the prior distribution hyperparameters of the Bayesian estimation. and Apply exponential decay, with a decay factor of 1. and Then, the new observation count within the current window is superimposed.

[0042] The probability estimates corresponding to the three sliding windows are weighted and fused together. The weights are adaptively adjusted according to the number of observed samples in each window. The window with a larger sample size receives a higher weight, thus obtaining the probability of normal occurrence.

[0043] Furthermore, the causal chain generation module: based on the determined causal relationship edges, performs bidirectional growth to obtain the original causal event chain, queries a preset physical constraint knowledge base to prune the original causal event chain, and outputs the structured causal event chain object. The physical constraint knowledge base stores the causal role attributes and causal pairing taboos of each weak electronic system. The causal role attributes represent that subsystem events can only be causes, only effects, or both. The causal pairing taboos are subsystem event association pairs that do not conform to physical causal logic.

[0044] The system also includes a closed-loop evolution module, which receives confirmation, negation, or correction feedback from maintenance personnel on the output causal event chain objects, converts the feedback results into virtual counts, and injects them in incremental form into the Bayesian estimator maintained by the normal probability maintenance module and the causal direction determination module, so that the probability estimation of the system is optimized as maintenance experience accumulates.

[0045] The technical effects and advantages of this invention are as follows: First, this invention is executed by edge computing nodes deployed in various production areas of the factory. These edge computing nodes acquire real-time anomalous signals generated by each weak electronic system. Once a signal is marked as a seed signal, a spatiotemporal correlation window is dynamically determined based on its subsystem type and anomalous amplitude. Anomalous signals from other subsystems falling within this window are then selected to form candidate causal pairs. This method enables the automatic discovery of correlations between anomalous signals across subsystems, eliminating the need for manual comparison.

[0046] Second, this invention maintains the normal occurrence probability for each type of abnormal event as a benchmark for causal inference. For each candidate causal pair, it calculates the causal lift in two directions, determines the causal relationship direction by comparing the bidirectional causal lift, obtains the causal relationship edges, and concatenates these edges into a causal event chain. After pruning through a physical constraint knowledge base, it outputs a structured object. This process transforms the causal discovery problem into probability lift calculation, eliminating the need for predefined causal rules or labeled samples, thus improving the automation level of cross-system causal relationship discovery.

[0047] Third, after reviewing the causal event chain objects, maintenance personnel submit confirmation, denial, or correction feedback. The edge computing nodes convert this feedback into virtual counts and inject them into the Bayesian estimator, incrementally updating the probability of normal occurrence and conditional probability, and writing repeatedly denied causal chain patterns into the physical constraint knowledge base. This closed-loop mechanism helps improve the accuracy of causal relationship determination. Attached Figure Description

[0048] Figure 1 This is a schematic diagram of the method control process of the present invention; Figure 2 This is a system module composition framework diagram of the present invention. Detailed Implementation

[0049] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. 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.

[0050] Example 1 As attached Figures 1 to 2 The method and system for integrated intelligent operation and maintenance management of low-voltage electrical systems across the entire oil and petrochemical plant, based on edge computing, are shown. This method is executed by edge computing nodes deployed in various production areas of the plant.

[0051] Edge computing nodes are deployed according to the production area.

[0052] For example, deploy an edge computing node in the oil refining unit area to connect to the area's video surveillance system and vibration fiber optic perimeter system; Deploy an edge computing node in the storage and transportation tank area and connect it to the area's fire alarm system, environmental monitoring system, and access control system; Along the long-distance pipeline corridor, an edge computing node is deployed every 5km to 8km to connect to the vibration fiber optic perimeter system and video monitoring system of that section.

[0053] Each edge computing node establishes a communication connection with the weak electronic systems in its area via the factory's industrial Ethernet.

[0054] The following provides a detailed explanation of the specific implementation process of this method.

[0055] Step S1 is: Edge computing nodes acquire real-time abnormal signals generated by various weak electronic systems.

[0056] The video surveillance system captures images and detects targets in real time within a preset warning area. When personnel or vehicles are detected entering the warning area, an abnormal signal is generated.

[0057] The sensing fiber of the vibration fiber perimeter system is laid along the pipe gallery and the wall. When the vibration amplitude of a certain fiber exceeds the normal range, an abnormal signal is generated.

[0058] The access control system generates an abnormal signal when it detects unauthorized entry or abnormal card swiping behavior.

[0059] When a smoke detector or heat detector is triggered, the fire alarm system generates an abnormal signal.

[0060] The environmental monitoring system generates an abnormal signal when the concentration of combustible or toxic gases exceeds the safety threshold.

[0061] As one implementation method, after each low-voltage electronic system completes anomaly detection internally, it encapsulates the anomaly information into a message and actively pushes it to the edge computing node. Each message contains four fields: the subsystem type field indicates the source of the anomaly, with values ​​for video surveillance system, vibration fiber optic perimeter system, access control system, fire alarm system, or environmental monitoring system; The abnormal feature value field records the specific quantitative value of the abnormality, such as the amplitude decibel value detected by the vibration fiber, the gas concentration value collected by the environmental monitoring system, and the confidence level of the cross-boundary target detected by the video surveillance system. The timestamp field records the precise moment the exception occurred; The logical space coordinate field identifies the spatial location where the exception occurred.

[0062] Edge computing nodes receive and parse these messages to obtain a unified format of abnormal signal streams.

[0063] There are two ways to determine the coordinates in logical space.

[0064] One approach is to pre-establish a unified three-dimensional spatial coordinate system during the factory construction phase. The origin is the southwest corner of the factory area, with the positive X-axis pointing east, the positive Y-axis pointing north, and the positive Z-axis perpendicular to the ground and pointing upwards. Each sensing terminal of the low-voltage electronic system is calibrated with a three-dimensional coordinate within this system. For example, if a vibration fiber optic sensing unit in the pipe gallery area is calibrated as coordinates (230, 185, 3), then the logical spatial coordinates when this sensing unit generates an abnormal signal are (230, 185, 3).

[0065] Another approach is to use the existing equipment tag system in the factory area. Each terminal is assigned a unique equipment tag number. The edge computing node maintains a mapping table between the equipment tag number and the three-dimensional spatial coordinates. When it is necessary to calculate the spatial distance, the equipment tag number is converted into a three-dimensional coordinate value by looking up the table.

[0066] Both of these methods enable spatiotemporal correlation analysis of anomalous signals from different weak electronic systems within a unified spatial metric framework.

[0067] Step S2 is as follows: The edge computing node continuously monitors the abnormal signal stream acquired in step S1 in real time and discovers signal combinations that may have causal relationships.

[0068] Each time an edge computing node receives a new abnormal signal, it extracts its abnormal feature value and compares it with the preset alarm threshold of the subsystem to which the signal belongs. Each subsystem has its own alarm threshold. For example, the amplitude alarm threshold of the vibration fiber optic perimeter system is set at 30dB, and the hydrogen sulfide concentration alarm threshold of the environmental monitoring system is set at 10ppm according to national standards.

[0069] When an abnormal characteristic value exceeds the alarm threshold, the signal is marked as a seed signal, triggering a candidate causal pair generation process. The seed signal represents an abnormal event that occurred in the plant area, and subsequent cross-system causal relationship searches are carried out centered on it.

[0070] For the labeled seed signal, the edge computing node determines the time span and spatial radius of the spatiotemporal correlation window, both of which are dynamically calculated based on the specific characteristics of the seed signal.

[0071] The edge computing node maintains a configuration mapping table from subsystem type to window base scale, and presets a base value for the time span for each type of weak electronic system. and the basic value of spatial radius .

[0072] In one exemplary configuration: Fire alarm system For 60s, It is 100m; Environmental monitoring system For 60s, It is 200m; Video surveillance system For 30 seconds It is 50m; Vibrating fiber perimeter system For 10s, It is 30m; Access control system For 5s, It is 10m.

[0073] The above values ​​are set based on the physical propagation characteristics of abnormal events monitored by different subsystems. Smoke diffusion in fire alarm systems and gas leakage in environmental monitoring systems are diffusion-type anomalies, which spread widely in all directions when they occur, affecting a wide area and requiring a certain amount of time to propagate. Therefore, a larger base window is configured.

[0074] Unauthorized intrusion into the access control system and localized vibration of the fiber optic perimeter system are localized anomalies. The anomaly itself only occurs at a specific location, and the impact is concentrated in the vicinity of the incident point. Therefore, a smaller base window is configured. The base scale of diffuse anomalies is generally larger than that of localized anomalies.

[0075] Building upon the base scale, edge computing nodes scale the window based on the severity of anomalies in the seed signal. The anomaly amplitude A is defined as the ratio of the anomaly characteristic value of the seed signal to its alarm threshold. For example, if the alarm threshold for a vibrating optical fiber is 30dB, and the actual detected vibration amplitude is 45dB, then A is 1.5.

[0076] The rule for determining the window scaling factor k is as follows: When A does not exceed 1, k takes the value 1; When A is greater than 1 ,in The preset scaling factor, The value ranges from 0.3 to 1.0.

[0077] At the same time, k has an upper limit m to prevent the window from expanding indefinitely. For example, take... If A is 0.5 and m is 5, then in the example above, when A is 1.5, .

[0078] The larger the anomalous amplitude, the larger the window scaling factor. The final calculated time span of the spatiotemporal correlation window... spatial radius .

[0079] After obtaining the spatiotemporal correlation window, the edge computing nodes filter candidate signals from the anomalous signal stream. The filtering criteria simultaneously satisfy the following three conditions: First, the timestamp of the signal falls within a region centered on the timestamp t of the seed signal, extending before and after it. Within the range; Second, the three-dimensional Euclidean distance between the logic space coordinates of the signal and the logic space coordinates of the seed signal is less than... ; Third, the type of the subsystem to which the signal belongs is different from the type of the subsystem to which the seed signal belongs.

[0080] The first two conditions ensure that the candidate signal and the seed signal are sufficiently close in time and space, while the third condition limits the consideration to cross-subsystem correlations. Each selected signal that meets the conditions is paired with the seed signal to form a candidate causal pair, where the first element is the seed signal and the second element is the selected signal.

[0081] Step S2 also includes an adaptive adjustment mechanism for adjusting the window scaling factor. Continuous optimization is performed. Edge computing nodes maintain a moving average of the number of candidate causal pairs for each type of anomalous event, recording the average number of candidate causal pairs generated each time when this type of event is used as a seed signal over a defined period of time. When the sliding average value of a certain type of event exceeds the preset upper limit for an extended period, the scaling factor corresponding to that type of event will be adjusted. Adjusted to 0.9 times the original value; When the moving average is zero for a long period of time, Adjusted to 1.1 times the original value, the adjusted value The value does not exceed the preset upper limit. This adjustment keeps the number of candidate causal pairs within the preset range.

[0082] Step S3 is: Edge computing nodes maintain the normal occurrence probability for each type of abnormal event in the low-voltage system, serving as the benchmark for causal inference in step S4.

[0083] An abnormal event category refers to a set of events of the same abnormal type generated by the same subsystem type. For example, "video surveillance system - area intrusion", "vibration fiber optic perimeter system - vibration abnormality", "access control system - illegal entry", "fire alarm system - smoke alarm", and "environmental monitoring system - gas concentration exceeding the standard" are each independent event categories, and the normal occurrence probability of each event category is maintained separately.

[0084] The system employs three sliding windows with different time granularities to work in parallel, and the probability estimates of each window are weighted and fused.

[0085] Short time windows cover the most recent 1 hour and are updated in 1-minute increments. Medium time windows cover the most recent 24 hours and are updated in 10-minute increments. Long time windows cover the most recent 7 days and are updated in 1-hour increments. Each window records only observations of this type of event occurring independently. Independent occurrence refers to a predetermined time range before and after the event's occurrence. No anomalous signals from other subsystems were found that could form candidate causal pairs with the event. For example, take 10 seconds.

[0086] Each sliding window employs incremental Bayesian estimation with a forgetting mechanism. Taking a short window as an example, the system maintains two hyperparameters for it. and The initial values ​​are all set to 1. An update operation is performed every 1 minute: first, the initial values ​​are set to 1. and Each multiplied by a decay factor , The value is greater than 0 and less than 1, for example, 0.9; then, new observation counts are added based on whether this type of event has occurred independently within this step size. If at least one independent occurrence has occurred, then... Increase by 1, or if it does not occur, then... Increase by 1. Update the probability estimate for this short time window. .

[0087] The medium and long time windows use the same update method, differing only in step size and decay factor: the medium time window has a step size of 10 minutes and a decay factor of... Set to 0.99; long time window step size 1h, decay factor Take 0.999.

[0088] Obtain the probability estimates for each of the three windows. , , Then, the edge computing nodes weight and fuse them to obtain the normal occurrence probability P. The fusion weight is determined by the effective sample size of each window, which is defined as the sum of the two hyperparameters of that window, i.e. , , Fusion weights for short time windows Weights of medium and long time windows , The same calculation applies. Probability of normal occurrence. .

[0089] In the initial stage of system startup, the effective sample size of short time windows is relatively large and the weight is high; as the running time increases, the weight of medium and long time windows increases.

[0090] Step S4 is: The edge computing node determines the direction and strength of the causal relationship for each candidate causal pair generated in step S2. The two events in the candidate causal pair are denoted as X and Y, respectively.

[0091] The system maintains a conditional probability Bayesian estimator for the X→Y direction, which operates in the same manner as the single-window Bayesian estimation in step S3. The hyperparameters of this estimator are denoted as... and The initial values ​​are all set to 1, and the attenuation factor is denoted as . and For example, take 0.9.

[0092] After each event X occurs, the system observes for a preset time window. Does the internal observation event Y occur subsequently? The value of covers the delay time required for a cause event to propagate to a result event in the physical world, for example, 10 seconds.

[0093] If event Y is in If it occurs within the time frame, then update. , ; If event Y is in If it has not occurred, then update. , .

[0094] The updated conditional probability estimate is obtained. Using the same approach, the system maintains hyperparameters in the Y→X direction. and Obtain conditional probability estimates .

[0095] The system calculates the causal lift in both directions. Causal lift in the X→Y direction. , representing the multiple of the probability of event Y occurring after event X occurs, relative to the normal probability of Y occurring. For example, if It is 0.1. If it is 0.6, then It is version 6.0.

[0096] Similarly, calculate the causal lift in the Y→X direction. .in and The normal occurrence probabilities of the event categories to which events X and Y belong are obtained from the results maintained in step S3. Then, the directional asymmetry ratio is calculated. .

[0097] The system internally presets three decision thresholds: an upper threshold U, a lower threshold D, and a confidence threshold C. In an exemplary configuration, U is set to 2.0, D is set to the reciprocal of U (0.5), and C is set to 3.0.

[0098] The judgment rule is as follows: when and When determining the direction of causality, X points to Y, with X as the cause and Y as the effect, and... The value is used as the confidence weight of the causal relationship edge; when and At that time, the direction of the causal relationship is determined to be Y pointing to X, in order to As a confidence weight; when When the condition is met, it is determined to be a symmetrical coupling relationship, and no causal relationship edge is generated. The upper threshold U is greater than 1 and the lower threshold D is the reciprocal of U.

[0099] Each causal relationship edge confirmed by the judgment obtains three attributes: causal event identifier, result event identifier, and confidence weight, and is included in the causal relationship edge set.

[0100] Step S5 is as follows: The edge computing node concatenates the set of causal relationship edges obtained in step S4 into a causal event chain, and outputs it after physical logic verification.

[0101] The causal event chain is constructed by starting from one edge and extending to both ends.

[0102] First, select any established causal relationship edge as the initial edge. When extending towards the front, search the edge set for other edges whose result is the current chain head event. If a matching edge exists, link it to the front of the chain, and use the cause event of the newly linked edge as the new chain head to continue searching forward until no more preceding relationships can be found.

[0103] When extending backward, the system searches the edge set for other edges whose cause is the current chain tail event. If a matching edge exists, it is linked to the next event in the chain, and the result event of the newly linked edge becomes the new chain tail, continuing the search backward until no more subsequent relationships are found. This forward and backward extension is performed alternately, resulting in an original causal event chain that cannot be extended further at either end.

[0104] For example, suppose the edge set contains two causal edges: “perimeter vibration → access control intrusion” and “access control intrusion → video out of bounds”. Taking “access control intrusion → video out of bounds” as the initial edge, when extending to the back end, “perimeter vibration → access control intrusion” is found and linked to the front of the chain, resulting in the complete causal chain “perimeter vibration → access control intrusion → video out of bounds”.

[0105] The system pre-defines a physical constraint knowledge base within the edge computing node, which consists of two data tables.

[0106] The first table is the causal role attribute table, recording the causal role that each subsystem event type can play in the physical world. There are three possible roles: cause only, effect only, and causal. This table was pre-determined by experts in the oil and petrochemical safety field based on the physical characteristics of each subsystem event. For example, a gas concentration exceeding the standard event in an environmental monitoring system is labeled as "effect only," a smoke alarm event in a fire alarm system is labeled as "effect only," and an unauthorized entry event in an access control system is labeled as "causal."

[0107] The second table is the causal pairing taboo table. Each record stores a prohibited causal pair, containing two fields: cause event type and result event type. For example, "fire alarm event → access control intrusion event" is listed as a taboo pair, and "gas concentration exceeding the standard event → perimeter vibration event" is listed as a taboo pair. The initial content of this table is preset by experts during system deployment and supplemented through feedback loop during operation.

[0108] The pruning process is as follows: The edge computing node traverses each edge of the original causal event chain, extracts the event types of the cause event and the result event of that edge, and queries the causal role attribute table to obtain the causal role of both.

[0109] If the causal role of the cause event is "can only be an effect", then remove that edge and all edges after it; if the causal role of the result event is "can only be a cause", then similarly remove that edge and all edges after it.

[0110] Then, the cause event type and result event type corresponding to the edge are combined into a type pair, which is then matched against each record in the causal pairing taboo table. If a match is successful, the edge is pruned. The remaining link fragments after physical and logical pruning constitute the output structured causal event chain object.

[0111] As one implementation method, the structured causal event chain object uses a JSON-encoded format and contains three fields: The root cause event identifier field indicates the starting event of the entire chain. The event node sequence field contains event nodes arranged in chronological order. Each node provides the subsystem type, exception type, timestamp, and logical space coordinates. The confidence weight sequence field provides the confidence weight values ​​for each causal relationship segment in the order of the causal relationship edges between event nodes. This object is pushed to the operation and maintenance monitoring terminal via the MQTT protocol, where the causal chain is displayed graphically on the terminal interface.

[0112] The feedback loop is as follows: After viewing the causal event chain object on the terminal, the operations and maintenance personnel can submit confirmation feedback, negative feedback, or correction feedback through the interface.

[0113] When maintenance personnel submit confirmation feedback, the edge computing node traverses each causal relationship edge on the confirmation chain, performs an occurrence update on the conditional probability Bayesian estimator corresponding to each edge's direction, i.e., updates the hyperparameters. Increase by 1; Meanwhile, for the outcome events of each edge on the chain, within the short time window of its normal occurrence probability estimator, if the event occurs at a preset time... If no candidate causal pair is formed between the internal signal and abnormal signals from other subsystems, no action will be taken; otherwise, [action will be taken]. Increase by 1.

[0114] When operations and maintenance personnel submit negative feedback, the edge computing node performs an update once in the conditional probability estimator corresponding to the direction of each edge on the chain, which means no update occurs. Increment by 1, and increment the negation count corresponding to the causal chain pattern by 1.

[0115] When operations and maintenance personnel submit correction feedback and specify the corrected causal direction and link, the edge computing node executes an update in the conditional probability estimator corresponding to the correct direction, but does not execute an update in the conditional probability estimator corresponding to the original incorrect direction.

[0116] If the number of times a certain causal event chain pattern receives negative feedback reaches a preset threshold, such as 3 times, the system writes the corresponding cause event type and result event type as a new record into the causal pairing taboo table of the physical constraint knowledge base. Subsequently, when the system encounters a causal chain with the same pattern again, it will be automatically removed during the pruning operation in step S5.

[0117] The system consists of the following components: The present invention also provides an intelligent operation and maintenance management system for low-voltage electrical systems in the entire oil and petrochemical plant based on edge computing. The system is deployed in the edge computing node of the plant and consists of a candidate causal pair generation module, a normal probability maintenance module, a causal direction determination module, a causal chain generation module, and a closed-loop evolution module.

[0118] The candidate causal pair generation module is used to execute the processing logic of steps S1 and S2. This module internally includes a data interface unit and a window calculation unit. (1) The data interface unit receives the abnormal signal message and parses it into subsystem type, abnormal feature value, timestamp and logical space coordinates.

[0119] (2) The window calculation unit stores a configuration mapping table between subsystem types and window base scales. When a signal is marked as a seed signal, the corresponding value is searched. and The process involves calculating the anomalous amplitude and window scaling factor, determining the spatiotemporal correlation window, filtering out anomalous signals that meet the criteria to form candidate causal pairs, and maintaining a moving average of the number of candidate causal pairs to apply the scaling factor. Adjustments were made based on feedback.

[0120] The normal probability maintenance module is used to execute the processing logic of step S3. This module internally includes a multi-window management unit and an attenuation control unit. (1) The multi-window management unit maintains three data structures for each type of abnormal event: short time window, medium time window, and long time window, and stores hyperparameters respectively. and , and , and .

[0121] (2) The attenuation control unit stores the attenuation factor. , and The update operation is triggered periodically according to the update cycle of each window, and the probability estimates of the three windows are weighted and merged into the normal occurrence probability.

[0122] The causal direction determination module is used to execute the processing logic of step S4. This module internally includes a conditional probability estimation unit and a direction decision unit. (1) The conditional probability estimation unit maintains a Bayesian estimator and its hyperparameters for each candidate causal pair in two directions, and updates them incrementally based on whether another event occurs within the observation window after the corresponding event occurs.

[0123] (2) The direction decision unit reads the normal occurrence probability of the relevant event from the normal probability maintenance module, calculates the causal lift degree and directional asymmetry ratio of the two directions, compares them with the preset threshold, and outputs the causal relationship edge with directional label and confidence weight.

[0124] The causal chain generation module is used to execute the processing logic of step S5: This module internally includes a bidirectional growth unit and a pruning unit. (1) The bidirectional growth unit performs forward and backward search links in the set of causal relationship edges to connect the edges into the original causal event chain.

[0125] (2) The pruning unit stores two data tables of the physical constraint knowledge base, checks the compliance of causal roles and the taboo of causal pairing on each edge of the original chain, and outputs a structured causal event chain object.

[0126] The closed-loop evolution module is used to receive feedback from the operation and maintenance terminal. Internally, this module includes a feedback receiving unit, a sample injection unit, and a knowledge base update unit. (1) The feedback receiving unit parses the confirmed, denied, or corrected feedback. The sample injection unit writes the corresponding virtual count increment into the Bayesian estimator of the normal probability maintenance module and the causal direction determination module according to the feedback type.

[0127] (2) The knowledge base update unit tracks the number of negations of each causal chain pattern, and writes the corresponding subsystem event association pair into the causal pairing taboo table of the physical constraint knowledge base when the threshold is reached.

[0128] The above modules have several optional implementation methods: In one implementation, all modules are executed by the central processing unit of the edge computing node, which executes program instructions stored in the memory, and the modules exchange data through memory sharing.

[0129] As another implementation, the candidate causal pair generation module, the normal probability maintenance module, the causal direction determination module, and the causal chain generation module are implemented by field-programmable gate array hardware logic circuits, while the closed-loop evolution module is implemented by an independent microcontroller executing firmware programs. The two parts exchange data through a board-level data bus.

[0130] The specific parameter values ​​mentioned in the above embodiments, such as the basic time span and spatial radius of each subsystem, the time length and update step size of each sliding window, the values ​​of each attenuation factor, the set values ​​of each decision threshold, and the threshold for the number of feedback rejections, are all illustrative examples. In actual deployment and application, these parameters can be flexibly adjusted and optimized according to factors such as the specific scale of the plant area, the actual characteristics of each weak electronic system, and the statistical analysis results of the plant's historical operating data. These adjustments do not exceed the protection scope defined by the claims of this invention.

Claims

1. A method for integrated intelligent operation and maintenance management of low-voltage electrical systems across an entire oil and petrochemical plant based on edge computing, characterized in that: Edge computing nodes deployed in the factory area perform the following steps: S1. Acquire abnormal signals generated in real time by multiple weak electronic systems within the oil and petrochemical plant area. The weak electronic systems include video surveillance system, vibration fiber optic perimeter system, access control system, fire alarm system, and environmental monitoring system. Each abnormal signal includes the type of its subsystem, abnormal feature value, timestamp, and logical space coordinates. S2. For any weak electrical system signal identified as an anomalous change at the current moment, take it as a seed signal. Determine the basic scale of the window according to the subsystem type to which the seed signal belongs, and determine the window scaling factor according to the anomalous amplitude of the seed signal. The time span and spatial radius of the spatiotemporal correlation window are obtained by multiplying the basic scale of the window and the window scaling factor. Anomalous signals from other weak electrical systems whose timestamps and logical spatial coordinates fall within the spatiotemporal correlation window are combined with the seed signal to form candidate causal pairs. S3. Maintain the normal occurrence probability for each type of weak current system abnormal event. This normal occurrence probability is based on the independent occurrence record of the abnormal event when it does not form a candidate causal pair with other weak current system abnormal signals, and is dynamically updated by incremental Bayesian estimation with forgetting. S4. For each candidate causal pair, calculate the causal lift in two directions. By comparing the causal lift in two directions, determine the causal relationship direction and obtain the causal relationship edge with direction label and confidence weight. The causal lift is the ratio of the conditional probability of an event after another event occurs to the normal probability of the event. S5. Connect multiple causal relationship edges in chronological order and causal direction to form a causal event chain. After pruning by preset physical logic constraints, output a structured causal event chain object. The causal event chain object includes root cause event identifiers, event node sequences, and confidence weights for each causal relationship edge.

2. The method for integrated intelligent operation and maintenance management of low-voltage electrical systems across the entire petroleum and petrochemical plant based on edge computing as described in claim 1, characterized in that, In step S2: The base window size for diffuse anomalies is larger than that for fixed-point anomalies. The abnormal amplitude is positively correlated with the window scaling factor.

3. The method for intelligent operation and maintenance management of low-voltage electrical systems across the entire petroleum and petrochemical plant based on edge computing as described in claim 2, characterized in that, Step S2 also includes: Monitor the number of candidate causal pairs generated by a single seed signal; When the number of candidate causal pairs exceeds the set upper limit, reduce the window scaling factor corresponding to this type of abnormal event; When the number of candidate causal pairs is 0, increase the window scaling factor corresponding to this type of abnormal event.

4. The method for intelligent operation and maintenance management of low-voltage electrical systems across the entire petroleum and petrochemical plant based on edge computing as described in claim 1, characterized in that, Step S3 includes: For each type of abnormal event in a low-voltage system, maintain three parallel, independent occurrence recording sliding windows: a short time window, a medium time window, and a long time window; For each independent occurrence record within a sliding window, the incremental Bayesian estimation with forgetting is used to obtain the probability estimate for that window, wherein each update first adjusts the prior distribution hyperparameters of the Bayesian estimation. and Apply exponential decay, with a decay factor of 1. and Then, the new observation count within the current window is superimposed; The probability estimates corresponding to the three sliding windows are weighted and fused together, and the weights are adaptively adjusted according to the number of observed samples in each window to obtain the probability of normal occurrence.

5. The method for intelligent operation and maintenance management of low-voltage electrical systems across the entire petroleum and petrochemical plant based on edge computing as described in claim 1, characterized in that, Step S4 includes: For the first event X and the second event Y in the candidate causal pair definition , ,in Let Y be the conditional probability of X after X occurs. Let Y be the probability of the normal occurrence. Let X be the conditional probability of Y after Y occurs. Let X be the normal occurrence probability; calculate ; when Greater than the preset upper threshold and When the confidence threshold is exceeded, the causal relationship is determined to be X pointing to Y, and then... As a confidence weight; when Less than the preset lower threshold and When the confidence threshold is greater than the given threshold, the causal relationship direction is determined to be Y pointing to X, and then... As a confidence weight; when When the value falls between the preset lower threshold and the preset upper threshold, it is determined to be a symmetrical coupling relationship.

6. The method for intelligent operation and maintenance management of low-voltage electrical systems across the entire petroleum and petrochemical plant based on edge computing as described in claim 1, characterized in that, Step S5 includes: Based on the determined causal relationship edges, search forward along the time axis for subsequent relationship edges that are caused by the current result event and perform forward growth. Search backward along the time axis for preceding relationship edges that are caused by the current result event and perform reverse tracing. Alternate between these steps until neither end of the link can be extended, thus obtaining the original causal event chain. The preset physical constraint knowledge base is queried, which stores the causal role attributes and causal pairing taboos of each weak electronic system. The causal role attributes represent that a subsystem event can only be a cause, only an effect, or both. The causal pairing taboos are subsystem event association pairs that do not conform to physical causal logic. When a link segment in the original causal event chain violates the causal role attribute or hits the causal pairing taboo, it is pruned, and the pruned causal event chain is output as the structured causal event chain object.

7. The method for intelligent operation and maintenance management of low-voltage electrical systems across the entire petroleum and petrochemical plant based on edge computing as described in claim 6, characterized in that, Also includes: Receive confirmation, denial, or correction feedback from operations and maintenance personnel regarding the output causal event chain objects; The feedback results are converted into virtual counts and injected incrementally into the Bayesian estimators corresponding to the normal occurrence probability and the conditional probability. The physical constraint knowledge base is updated based on the repeatedly denied causal event chain pattern.

8. A smart operation and maintenance management system for low-voltage electrical systems across an entire oil and petrochemical plant based on edge computing, characterized in that: This includes the following modules deployed within the factory's edge computing nodes: The candidate causal pair generation module is configured to receive real-time abnormal signals from multiple weak electronic systems within the oil and petrochemical plant area. These weak electronic systems include video surveillance systems, vibration fiber optic perimeter systems, access control systems, fire alarm systems, and environmental monitoring systems. The module identifies abnormally abrupt weak electronic system signals as seed signals, determines the basic window scale based on the subsystem type to which the seed signal belongs, determines the window scaling factor based on the abnormal amplitude of the seed signal, and obtains the time span and spatial radius of the spatiotemporal correlation window by multiplying the basic window scale and the window scaling factor. The module then filters out other subsystem abnormal signals whose timestamps and logical spatial coordinates fall within the spatiotemporal correlation window and forms candidate causal pairs with the seed signals. The normal probability maintenance module is configured to collect independent occurrence records of each type of abnormal event in the low-voltage system when no candidate causal pair is formed, and update the normal occurrence probability using incremental Bayesian estimation with forgetting. The causal direction determination module is configured to calculate the causal lift in both directions for each candidate causal pair, determine the causal relationship direction by comparing the bidirectional causal lift, and output the causal relationship edge with direction label and confidence weight. The causal lift is the ratio of the conditional probability of an event after another event occurs to the normal probability of the event. The causal chain generation module is configured to concatenate a set of causal relationship edges into a causal event chain according to time sequence and causal direction. It calls a preset physical constraint knowledge base to prune the concatenation result and outputs a structured causal event chain object, which contains the root cause event identifier, the event node sequence, and the confidence weight of each causal relationship edge.

9. The intelligent operation and maintenance management system for low-voltage electrical systems across the entire petroleum and petrochemical plant based on edge computing as described in claim 8, characterized in that, The normal probability maintenance module is specifically configured as follows: For each type of abnormal event in a low-voltage system, maintain three parallel, independent occurrence recording sliding windows: a short time window, a medium time window, and a long time window; For each independent occurrence record within a sliding window, the incremental Bayesian estimation with forgetting is used to obtain the probability estimate for that window, wherein each update first adjusts the prior distribution hyperparameters of the Bayesian estimation. and Apply exponential decay, with a decay factor of 1. and Then, the new observation count within the current window is superimposed; The probability estimates corresponding to the three sliding windows are weighted and fused together, and the weights are adaptively adjusted according to the number of observed samples in each window to obtain the probability of normal occurrence.

10. The intelligent operation and maintenance management system for low-voltage electrical systems across the entire petroleum and petrochemical plant based on edge computing as described in claim 8, characterized in that, The causal chain generation module is specifically configured as follows: Based on the determined causal relationship edges, a bidirectional growth is performed to obtain the original causal event chain. The original causal event chain is pruned by querying a preset physical constraint knowledge base, and the structured causal event chain object is output. The physical constraint knowledge base stores the causal role attributes and causal pairing taboos of each weak electronic system. The causal role attributes represent that the subsystem event can only be a cause, can only be an effect, or can be both cause and effect. The causal pairing taboos are subsystem event association pairs that do not conform to physical causal logic. The system also includes a closed-loop evolution module, which is configured to receive confirmation, negation, or correction feedback from maintenance personnel on the output causal event chain objects, convert the feedback results into virtual counts, and inject them in incremental form into the Bayesian estimator maintained by the normal probability maintenance module and the causal direction determination module.

Images