An oilfield station alarm root cause analysis and disposal method based on a knowledge graph

Abstract

The application relates to the field of oilfield alarm analysis, and discloses an oilfield station alarm root cause analysis and disposal method based on a knowledge graph, which comprises the following steps: standardizing processing of multi-source operation data and calculating confidence of a measuring point to generate a standard alarm event; constructing a root cause constraint graph based on a process flow and an equipment account book, combining alarm characteristics to search and generate a candidate cause set; screening an effective process path for the candidate cause set, calculating a propagation delay, and generating a dynamic verification window of multi-source evidence; extracting a time sequence data segment in the dynamic verification window to confirm characteristics and calculate a comprehensive evidence score; determining a main cause according to the score and a conflict rejection rule, deducing an influence range along the root cause constraint graph, and determining a risk level; generating a disposal suggestion based on the risk level and issuing a control instruction; and finally, performing controlled knowledge updating on graph parameters according to a recovery verification result. The application effectively improves alarm root cause positioning accuracy and automatic disposal capacity.

Description

Technical Field

[0001] This invention relates to the field of oilfield alarm analysis, specifically a method for root cause analysis and handling of oilfield station alarms based on knowledge graphs. Background Technology

[0002] Oilfield stations generate a large amount of monitoring data during operation, and accurate analysis of alarm root causes is crucial for ensuring safe production. Existing alarm analysis methods for oilfield stations typically directly call upon underlying sensor data for logical judgments, lacking assessment of data source quality and fault tolerance mechanisms. When a single sensor fails or underlying data fluctuates abnormally, the system cannot supplement the data, easily leading to deviations in subsequent analysis results.

[0003] Meanwhile, the station's process pipeline network has a complex physical topology, and changes in data status experience physical propagation delays between different nodes. Existing methods often use fixed time windows for matching when fusing multi-source data, failing to fully consider the dynamic time differences brought about by the process path. This can easily lead to time matching deviations under complex operating conditions, making it difficult to ensure the temporal accuracy of evidence extraction.

[0004] Due to the lack of a cross-validation mechanism based on multi-source evidence, traditional methods rely heavily on human experience when investigating primary causes, making it difficult to effectively eliminate interference from conflicting data. These methods generally lack a self-feedback verification process based on on-site handling and recovery, and cannot correct the judgment logic through closed-loop iteration, thus limiting the accuracy of root cause localization and the system's applicability in long-term operation. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides a knowledge graph-based root cause analysis and handling method for oilfield station alarms. This method solves the problems of existing oilfield station alarm analysis methods relying on manual experience or a single fixed threshold, difficulty in multi-source data fusion under complex operating conditions, low root cause location accuracy, and lack of automated verification and closed-loop handling.

[0006] To achieve the above objectives, the present invention provides the following technical solution: a knowledge graph-based root cause analysis and handling method for oilfield station alarms, comprising the following steps:

[0007] Perform standardization processing on multi-source operational data to output standardized time-series data, calculate the measurement point confidence of the standardized time-series data, and generate standard alarm events containing alarm features;

[0008] A root cause constraint map is constructed based on process flow and equipment ledger information, and a candidate cause set is generated by retrieving the root cause constraint map using the alarm features;

[0009] Valid process paths are selected from the candidate cause set, and the propagation delay of the valid process paths is calculated in combination with the confidence of the measurement points; multi-source evidence is obtained, the causal time direction and reference event time of the multi-source evidence are determined, and a dynamic verification window of the multi-source evidence is generated based on the causal time direction, the propagation delay of the valid process paths and the reference event time.

[0010] The execution features of the time-series data segments corresponding to the multi-source evidence are extracted within the dynamic verification window for verification, and a comprehensive evidence score is calculated.

[0011] Based on the comprehensive evidence score and the pre-set conflict and exclusion rules in the root cause constraint map, the candidate cause set is classified into definite causes, and the scope of influence is deduced along the root cause constraint map to determine the event risk level;

[0012] Based on the event risk level, a handling suggestion and control instruction are generated, the control instruction is issued to execute the handling, recovery verification is performed, and the graph parameters in the root cause constraint graph are updated with controlled knowledge.

[0013] This invention combines station topology with multi-source operational data. It assesses data source quality by calculating the confidence level of measurement points in time-series data, providing a data foundation for subsequent judgments. Process flow attributes are transformed into search constraints for the knowledge graph, narrowing the scope of root cause investigation. A dynamic verification window is generated by calculating the propagation delay of effective process paths, aligning the temporal relationships of multi-source evidence to reduce the impact of operational fluctuations on time matching. Cross-validation using positive scores, negative penalty scores, and conflict exclusion rules reduces the misjudgment rate caused by single-point sensor failures. Finally, controlled knowledge updates are performed on the knowledge graph parameters using operational recovery results, achieving parameter iteration of the knowledge graph and improving the accuracy of system operation.

[0014] Furthermore, the multi-source operational data undergoes standardization processing to output standardized time-series data. The measurement point confidence level of the standardized time-series data is calculated, and a standard alarm event containing alarm characteristics is generated, including:

[0015] Acquire the multi-source operational data and establish a unified time scale axis for the sampling sequences of different data source channels in the multi-source operational data;

[0016] Perform a dimensional normalization operation to convert the engineering units extracted from different data source channels into a unified standard physical unit;

[0017] Perform a dead value marking operation to mark the sampled sequence as a frozen dead value, and complete the standardization process to obtain the standardized time series data;

[0018] Extract the quality code and time series stability from the standardized time series data, and calculate the measurement point confidence level corresponding to the data source channel based on the extracted quality code and time series stability.

[0019] Based on the standardized time-series data, the monitoring conditions generate the standard alarm event, which includes the trigger time and the alarm characteristics.

[0020] Furthermore, a root cause constraint map is constructed based on the process flow and equipment ledger information. The alarm features are then used to retrieve and generate a candidate cause set from the root cause constraint map, including:

[0021] Based on the process flow and equipment ledger information, construct the node set of station equipment and pipeline processes required for the root cause constraint graph, and establish relationship edges between the node sets to associate different entities;

[0022] Write the edge validity condition for the relationship edge;

[0023] Extract the alarm features contained in the standard alarm event, and map the triggering device and triggering point in the alarm features as initial anchor points to the entity nodes corresponding to the root cause constraint map;

[0024] Using the initial anchor point as the starting point, the search logic is executed by comparing the edge validity conditions of the relation edge with the current running mode of the standard alarm event record. Relationship edges that do not meet the edge validity conditions are skipped, and the candidate cause nodes mapped to the end point of the search path are collected to generate the candidate cause set.

[0025] Furthermore, the process of selecting effective process paths from the candidate cause set includes:

[0026] Based on the steady-state equipment on / off status before the occurrence of the standard alarm event indicated in the standardized time series data, process paths connected to the candidate cause set are extracted from the root cause constraint map.

[0027] For the extracted process paths, paths that do not conform to the current operating mode are discarded to obtain the valid process paths.

[0028] Furthermore, calculating the propagation delay of the effective process path based on the confidence level of the measurement points includes:

[0029] Extract the real-time parameter data corresponding to the effective process path from the standardized time series data;

[0030] When the confidence level of the measurement point corresponding to the real-time parameter data is lower than the preset substitution trigger threshold, the associated substitution parameter data is retrieved from the root cause constraint map to generate a substitution estimate.

[0031] The real-time parameter data or the alternative estimated value is selected based on the confidence level of the measurement point.

[0032] For each relation edge in the effective process path, the corresponding delay time is determined based on the selected real-time parameter data or the alternative estimated value. The delay times corresponding to all relation edges included in the effective process path are numerically summed to obtain the propagation delay of the effective process path.

[0033] Furthermore, acquiring multi-source evidence, determining the causal time direction and reference event time of the multi-source evidence, and generating a dynamic verification window for the multi-source evidence based on the causal time direction, the propagation delay of the effective process path, and the reference event time, including:

[0034] Determine the causal time direction and reference event time corresponding to the multi-source evidence;

[0035] Based on the causal time direction, the propagation delay of the effective process path, and the reference event time, and by introducing time tolerance, the lower time limit and the upper time limit of the dynamic verification window are determined.

[0036] The dynamic verification window corresponding to the multi-source evidence is generated based on the lower time limit and the upper time limit.

[0037] Furthermore, within the dynamic verification window, the execution features of the time-series data segments corresponding to the multi-source evidence are extracted for verification, and a comprehensive evidence score is calculated, including:

[0038] Extract the time-series data segment corresponding to the multi-source evidence whose timestamp falls within the dynamic verification window;

[0039] The feature data of the time series data segment is extracted and logically verified with the preset constraint attribute data in the root cause constraint map. Based on the verification results, the feature matching state value and conflict intensity value corresponding to the multi-source evidence are determined.

[0040] The multi-source evidence is divided into positive evidence and conflicting evidence. A weight coefficient is assigned to the positive evidence, and an independent conflict weight coefficient is configured for the conflicting evidence.

[0041] A positive score is calculated based on the weight coefficient, feature matching state value, and corresponding test point confidence of a single positive piece of evidence; a negative penalty score is calculated based on the conflict weight coefficient, conflict intensity value, and corresponding test point confidence of a single conflict piece of evidence.

[0042] The comprehensive evidence score is calculated based on all the positive scores and all the negative penalty scores.

[0043] Furthermore, based on the comprehensive evidence score and the pre-defined conflict and exclusion rules in the root cause constraint map, the candidate cause set is classified into definitive causes, including:

[0044] The comprehensive evidence score is compared with a preset threshold for determining the cause.

[0045] If the comprehensive evidence score meets the determination threshold for determining the cause, and the conflict exclusion rule determines that there is no strongly conflicting evidence, then the corresponding candidate cause in the candidate cause set is marked as the determined cause.

[0046] If the strong conflict evidence is determined to exist according to the conflict exclusion rule, the corresponding candidate cause is marked as an excluded cause and removed from the candidate cause set;

[0047] If the comprehensive evidence score does not meet the determination threshold for cause and there is no strong conflicting evidence, then the corresponding candidate cause will be marked as an unmonitorable cause.

[0048] Furthermore, extrapolating the scope of influence and determining the event risk level along the root cause constraint map includes:

[0049] The primary root cause is determined by hierarchical ranking based on the causal topological location of the identified cause in the effective process path, the time of cause occurrence, the comprehensive evidence score, and the number of accompanying anomalies that can be explained.

[0050] Taking the nodes corresponding to the main root causes as the starting point of the deduction, a graph traversal expansion deduction is performed in the root cause constraint graph, and the deduction influence range is generated based on the effective traversal range.

[0051] The initial alarm level, process parameter deviation, duration of abnormal state, and importance of the main root cause-related equipment are obtained from the standard alarm events and the root cause constraint map.

[0052] A deterministic score is calculated based on the comprehensive evidence score and the duration of the abnormal state, and the severity is calculated based on the initial alarm level, the degree of deviation of the process parameters, the importance of the main root cause-related equipment, and the inferred impact range.

[0053] The event risk level is generated by matching the deterministic score with the severity input into a preset risk assessment matrix.

[0054] Furthermore, based on the event risk level, handling suggestions and control instructions are generated, the control instructions are issued to execute the handling, recovery verification is performed, and controlled knowledge updates are performed on the graph parameters within the root cause constraint graph, including:

[0055] Receive the event risk level and generate the handling suggestion and the control instruction, and perform a safety constraint verification on the control instruction based on the root cause constraint map;

[0056] When the safety constraint verification result indicates that execution is permitted, the control command is sent to the industrial control system;

[0057] The recovery verification is performed by tracking the recovery status of the operating conditions after the control command is issued.

[0058] Based on the recovery effect data obtained from the recovery verification and the accuracy evaluation of the root cause determination, the controlled knowledge update is performed on the graph parameters within the root cause constraint graph.

[0059] This invention provides a knowledge graph-based method for root cause analysis and handling of oilfield station alarms. It has the following beneficial effects:

[0060] 1. This invention establishes a data quality assessment and fault-tolerant supplementation mechanism by calculating the confidence level of measurement points in standardized time-series data and retrieving associated data from the root cause constraint map to generate alternative estimated values ​​when the confidence level is lower than the threshold. This effectively avoids interference to the analysis results caused by single sensor failures or abnormal fluctuations in underlying data, and improves the reliability of the data source and the accuracy of subsequent logical judgments.

[0061] 2. This invention calculates the propagation delay of the effective process path by combining the confidence level of the measurement points, and generates a dynamic verification window for multi-source evidence based on the causal time direction. The time series data segments are extracted within the dynamic verification window to perform feature confirmation, which solves the problem of time matching deviation caused by multi-source data fusion under complex working conditions and ensures the time series accuracy of evidence extraction.

[0062] 3. This invention calculates a comprehensive evidence score by combining positive scores and negative penalty scores, excludes strongly conflicting evidence based on conflict exclusion rules to determine the main root cause, and updates the spectrum parameters with controlled knowledge using the recovered verification data after the action is taken. This achieves cross-logical verification of multi-source evidence and closed-loop iteration of spectrum data, reduces the reliance of alarm analysis on human experience, and improves the accuracy of root cause localization and the long-term applicability of the system. Attached Figure Description

[0063] Figure 1 This is a system framework diagram of the present invention;

[0064] Figure 2 This is a flowchart of the method of the present invention;

[0065] Figure 3 This is a flowchart of the root cause constraint map construction and retrieval process of the present invention;

[0066] Figure 4This is a flowchart illustrating the dynamic process path identification and verification window deduction process of the present invention.

[0067] Figure 5 This is a flowchart illustrating the three-state determination and risk deduction process for the causes of this invention.

[0068] Figure 6 This is a flowchart illustrating the security gating verification and knowledge closed-loop process of the present invention.

[0069] Figure 7 The waveform comparison diagram is for the time-series data feature extraction based on the dynamic verification window of the present invention.

[0070] Figure 8 This is a comparison chart of multi-dimensional comprehensive evidence scores for candidate causes of the present invention. Detailed Implementation

[0071] The technical solutions in 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.

[0072] Reference Figure 1 This invention provides a knowledge graph-based oilfield station alarm root cause analysis and handling system, including a data acquisition module, a processing module, a reasoning module, and a decision-making module.

[0073] The data acquisition module obtains field instrument data, equipment operation data, and operation log data from the existing industrial control system of the oilfield station and the production management, equipment management, or operation and maintenance data sources associated with the industrial control system, and uses the acquired data as multi-source operation data. The industrial control system includes programmable logic controllers, remote terminal units, data acquisition and monitoring control systems, distributed control systems, and safety instrument systems. The production management, equipment management, or operation and maintenance data sources are used to provide equipment logs, manual operation records, inspection records, maintenance records, and historical operation data.

[0074] The processing module receives multi-source operational data acquired by the acquisition module, performs time axis alignment, dimension normalization, and dead value marking operations to evaluate the quality code and timing stability of the data channel, and then outputs standardized timing data and measurement point confidence.

[0075] The reasoning module constructs and internally stores a root cause constraint map based on the acquired process flow and equipment ledger information of the oilfield station. Based on the standardized time series data and measurement point confidence output by the processing module, it performs effective process path screening, propagation delay calculation, dynamic verification window generation, dynamic evidence retrieval, and cause three-state determination.

[0076] The decision module generates the final alarm root cause analysis and handling results and control commands based on the three-state determination results of the reasoning module. After performing safety constraint verification, the control commands are sent to the industrial control system. At the same time, recovery verification is performed and the graph parameters in the root cause constraint graph are updated with controlled knowledge. The safety constraint verification includes operator permission verification and mutual exclusion relationship verification of the current operating status of the equipment.

[0077] Reference Figure 2 This invention provides a knowledge graph-based method for root cause analysis and handling of oilfield station alarms, comprising the following steps:

[0078] S10. Multi-source operational data standardization, confidence evaluation, and alarm event construction. The processing module performs standardization processing on the multi-source operational data acquired by the acquisition module to output standardized time-series data. The standardization processing includes time axis alignment, dimension normalization, and dead value marking operations. Based on the quality code and time-series stability of the data channels, the module calculates the confidence of measurement points in different data source channels in the standardized time-series data. Simultaneously, based on the standardized time-series data monitoring conditions, when abnormal states such as threshold exceeding or state abrupt changes are detected in the standardized time-series data, the processing module generates a standard alarm event containing the trigger time and alarm characteristics.

[0079] S20. Construct a root cause constraint map and retrieve candidate causes. The inference module constructs a root cause constraint map based on the acquired process flow and equipment ledger information of the oilfield station. The root cause constraint map includes station equipment and pipeline processes characterizing the oilfield station's process flow, as well as physical constraint parameters used to calculate propagation delay. Using alarm features from standard alarm events, the inference module retrieves associated process nodes in the root cause constraint map and generates a set of candidate causes.

[0080] S30. Determine the current effective process path and calculate the dynamic verification window. For the candidate cause set, the inference module filters connected effective process paths from the root cause constraint map based on the current equipment on / off status indicated by standardized time series data. Combining real-time parameter data extracted from the standardized time series data or alternative estimates generated based on measurement point confidence, and utilizing the pre-set physical constraint parameters in the root cause constraint map, including spatial volume and thermal inertia parameters, the inference module calculates the propagation delay of the effective process path, obtains multi-source evidence, and generates a dynamic verification window for the propagation of causal evidence by combining the causal time direction of the multi-source evidence with the reference event time. The multi-source evidence includes key evidence, supporting evidence, and conflicting evidence.

[0081] S40. Dynamic evidence retrieval, evidence readiness determination, and comprehensive evidence calculation. Within the dynamic verification window, the inference module extracts time-series data segments corresponding to multi-source evidence used to verify the candidate cause set from standardized time-series data. When the current system time has not yet reached the time limit of the corresponding dynamic verification window, the multi-source evidence that still needs to be observed remains in an in-ready state; when the readiness conditions of the corresponding evidence are met, feature confirmation and comprehensive evidence scoring are performed based on the complete and valid time-series data segments.

[0082] S50. Cause tri-state determination, root cause ranking, impact scope deduction, and risk assessment. Based on the comprehensive evidence score and the pre-defined conflict and exclusion rules in the root cause constraint map, the inference module classifies the candidate cause set into identifiable causes, excluded causes, and unmonitorable causes. Excluded causes are removed, and unmonitorable causes are recorded and the current verification is terminated. For identifiable causes, the inference module sorts them according to their position in the effective process path to isolate accompanying anomalies, deduces the affected area along the root cause constraint map, and determines the event risk level.

[0083] S60. Handling Safety Constraints, Recovery Verification, and Controlled Knowledge Update. The decision module receives the event risk level and generates handling suggestions and control commands. It performs safety constraint verification on the control commands. After successful verification, the control commands are sent to the industrial control system in the form of request signals. The module also tracks the recovery status of the operating conditions after the control commands are issued to perform recovery verification. Subsequently, the controlled knowledge of the graph parameters in the root cause constraint graph is updated. The industrial control system executes the final equipment action according to its internal original hardware interlocking logic. The overall execution process does not change the safety interlocking protection settings of the industrial control system. The final alarm root cause analysis and handling result is output, forming a closed loop of alarm handling and knowledge update.

[0084] The method provided by this invention will be described in detail below with reference to the specific content of each step.

[0085] The standardization of multi-source operational data, confidence evaluation, and alarm event construction in step S10 include the following sub-steps:

[0086] S11. Standardized processing of multi-source operational data. The acquisition module reads multi-source operational data from the industrial control system and related production management, equipment management, or maintenance data sources; real-time location data is accessed in the form of raw location data streams, and equipment ledgers, inspection records, maintenance records, and historical operational data are accessed according to the timestamps or valid time intervals of the corresponding records.

[0087] The processing module receives multi-source operational data and raw point data streams transmitted from the acquisition module. Addressing the differences in sampling periods among data source channels due to different industrial communication protocols, it performs point mapping and time axis alignment operations. A unified time scale axis is established for the sampling sequences of different channels at fixed time intervals. For continuous analog quantity channels such as pressure, temperature, and flow rate, linear interpolation is used to supplement missing nodes on the unified time scale when the interval between adjacent valid sampling points does not exceed the preset maximum interpolation interval. The preset maximum interpolation interval is determined based on a comprehensive calculation of the physical inertia time constant of the process parameters and the maximum period of historical normal fluctuations.

[0088] For discrete state channels such as valve status, pump operation status, and switch feedback, a step-wise approach is used to maintain the previous valid state for time alignment. For event-type data such as alarm records, manual operation records, inspection records, and maintenance records, numerical interpolation is not performed. Instead, the original event timestamps are retained and mapped to the corresponding time positions on a unified time axis.

[0089] Missing data exceeding the maximum interpolation interval remains in a missing state and is marked as missing, preventing the use of interpolated data to span long data interruption intervals. Missing value identification is completed, followed by transition identification and quality code processing. Furthermore, the processing module performs dimensional normalization based on a pre-configured engineering mapping table, converting engineering units extracted from different data source channels into unified standard physical units.

[0090] After dimensional normalization, the processing module performs dead value marking by calculating the numerical variance of each data source channel within the sliding time window. When the numerical variance remains zero or its variation is below the dead zone threshold, and the oilfield station process is in dynamic operation, the processing module marks the sampled sequence as a frozen dead value and sets it as an invalid data segment unusable for normal numerical calculations. The standardized time series data retains the data source channel, the original values, and the frozen dead value marker, thus enabling subsequent steps to identify unusable data states while avoiding the loss of original records, and accordingly performing determinations based on reduced measurement point confidence, limited substitution, or inability to monitor. The dead zone threshold is set based on the instrument range accuracy and historical background noise level, ranging from 0.5% to 2% of the instrument range accuracy, to output standardized time series data. Time axis alignment interpolation and data structure transformation of multidimensional heterogeneous data sequences are performed using conventional methods.

[0091] S12. Calculate the confidence level of measurement points based on multi-dimensional evaluation rules. In actual operating conditions, process parameters at the same oilfield station are measured jointly by different data source channels or monitored jointly by primary and backup redundant instruments. The processing module calculates the confidence level of measurement points based on multi-dimensional evaluation rules. These rules weight six characteristic parameters related to data quality, temporal stability, and physical consistency. The calculation logic is expressed as follows:

[0092] ;

[0093] in, Indicates the first Confidence of measurement points for each data source channel; It represents the communication quality code score, reflecting the integrity of the underlying industrial protocol data packet transmission; The sampling coverage score reflects the density of the distribution of valid sampling points within the statistical period. The historical stability score reflects the inverse ratio of the frequency of glitches and noise occurring in the data source channel. The score indicates the consistency between the primary and backup points, reflecting the inverse ratio of the numerical deviations between two redundant instruments at the same process location; The process consistency score reflects the degree of fit between the data source channel values ​​and the media balance equations of the related upstream and downstream equipment. The score represents the collection time and reflects the freshness of the data timestamp and is inversely proportional to the delay in data synchronization and updating.

[0094] , , , , as well as This represents the corresponding weighting coefficient. Each weighting coefficient is obtained through regression analysis of historical instrument failure sample data, or determined using the analytic hierarchy process (AHP) based on field expert experience. Communication quality code scores, sampling coverage scores, historical stability scores, primary / backup point consistency scores, process consistency scores, and acquisition time scores are all converted to a unified scoring range of 0 to 1 before participating in the weighted calculation. Higher values ​​indicate higher data reliability for the corresponding dimension. All weighting coefficients are positive real numbers and have been normalized so that the sum of all weighting coefficients equals 1. Therefore, the measurement point confidence level... The value range is from 0 to 1.

[0095] Confidence level of measurement points obtained by weighted calculation The value range is constrained to be between 0 and 1. The processing module compares the measurement point confidence level with the internally set low-confidence marking threshold, which is calibrated based on the equipment's factory baseline accuracy and historical maintenance records, and ranges from 0.60 to 0.85. When the measurement point confidence level is lower than the low-confidence marking threshold, the corresponding data source channel and data are not deleted; instead, the data source channel is marked as a low-confidence point in the standardized time series data. Data source channels marked with low-confidence points are given lower priority in subsequent propagation parameter extraction, evidence matching, and security constraint verification calculations for graph inference, thereby preventing inferior data from dominating causal inference.

[0096] S13. Monitor abnormal operating conditions and establish standard alarm events. The processing module uses the standardized time-series data after cleaning to perform sliding window detection to extract process deviations. The triggering conditions involved in sliding window detection include threshold exceeding, state abrupt change, interlocking anomaly, and manually reported anomaly. Threshold exceeding corresponds to analog parameters exceeding the upper and lower safety limits set by the equipment; state abrupt change corresponds to unplanned jumps in the equipment start / stop feedback state without external control command intervention; interlocking anomaly corresponds to protection action signals triggered by the hardware logic of the safety instrumented system; and manually reported anomaly corresponds to unplanned state deviations entered by on-site operators through the terminal.

[0097] When any trigger condition is detected, the processing module obtains the absolute timestamp of the trigger action as a reference time and extracts associated attributes to generate a standard alarm event containing the trigger time and alarm characteristics. These alarm characteristics include alarm location identifier, alarm device identifier, alarm type, current operating mode, direction of operating condition change, magnitude of operating condition change, associated devices, and time-series data indexes before and after the alarm. The processing module uses the standard alarm event as the initial retrieval source and inputs it into the root cause constraint graph to initiate subsequent causal tracing reasoning. The sliding window detection process based on time-series data is implemented using conventional programming logic.

[0098] Reference Figure 3 The construction of the root cause constraint map and retrieval of candidate causes in step S20 includes the following sub-steps:

[0099] S21. Define and instantiate the entities and relation edges of the root cause constraint graph. The inference module reads the process flow drawings and equipment ledger information of the oilfield station, constructs the root cause constraint graph using the graph database, and establishes a set of nodes in the root cause constraint graph that covers the station equipment and pipeline processes representing the oilfield station process. The node set specifically includes entity types such as station, equipment, pipeline, valve, pump, instrument, alarm point, control point, feedback point, candidate cause, disposal measures, interlock conditions, and recovery criteria.

[0100] For each candidate cause node in the established set of nodes, the inference module inputs constraint attribute data including applicable alarm type, required location list, key evidence requirements, supporting evidence requirements, exclusion conditions, conflict conditions, direction of change, magnitude of change, duration, handling conditions, prohibition conditions, and recovery criteria. Simultaneously, the inference module establishes relationship edges between different entities within the node set. These relationship edges include equipment ownership relationships, upstream and downstream connection relationships, pipeline connection relationships, process influence relationships, cause-cause relationships, cause-elimination relationships, applicable handling relationships, and interlocking constraint relationships.

[0101] When constructing the root cause constraint graph, actual individual devices or instruments are mapped as node entities with attribute labels, and physical pipelines connecting different devices, control logic links, and causal inference rules are mapped as directional relationship edges. The operation of establishing nodes and relationship edges through the graph database is performed using conventional knowledge graph construction tools.

[0102] S22. Bind physical constraint parameters to graph relationship edges. To support dynamic verification window inference calculations, the inference module writes physical constraint parameters to the relationship edges in the root cause constraint graph that involve the propagation characteristics of physical quantities. The physical constraint parameters include the process propagation direction, delay model type, parameter points required for delay calculation, edge validity conditions, and allowed operation modes for propagation.

[0103] The process propagation direction indicates the propagation vector of the fluid medium or pressure wave in the pipeline. The delay model type distinguishes between pure time-delay delay due to medium transport and thermal inertia delay due to equipment dynamic response. The parameter points required for delay calculation point to standardized time-series data channel identifiers used to extract real-time flow or velocity data. Edge validity conditions define the physical prerequisites that must be met for an edge to be in a connected state, covering valve opening / closing status, pump operating status, medium flow direction conditions, and equipment start / stop status. Allowed operating modes for propagation define the applicability of the edge under different production conditions. The inference module attaches physical constraint parameters as attribute data structures to the corresponding edges, thereby forming a root cause constraint graph that integrates physical and process constraint attributes.

[0104] S23. Generate a candidate cause set based on standard alarm event retrieval. After receiving the standard alarm events output by the processing module, the inference module extracts alarm features such as alarm type, trigger point, triggering device, current operating mode, and upstream and downstream process relationships contained in the standard alarm events, and maps the triggering device and trigger point as initial anchor points to the entity nodes corresponding to the root cause constraint map.

[0105] Starting from the initial anchor point, the inference module uses a breadth-first search or depth-first search algorithm to perform search logic along upstream and downstream connections and causal relationships. During the search, the inference module compares the edge validity conditions of relation edges with the current operating mode recorded in the standard alarm event log, skipping relation edges that do not meet the edge validity conditions to narrow the search range. It also records visited entity nodes and terminates the search of the current branch when a visited node is encountered or the search depth reaches a preset depth limit threshold to avoid getting stuck in an infinite loop in a graph with a circular process topology. The depth limit threshold is determined based on the average number of process connection layers of adjacent key equipment within the oilfield station, typically ranging from 3 to 5 layers.

[0106] After the search logic is completed, the inference module obtains the candidate cause nodes mapped to the end point of the search path, and summarizes the matching candidate cause nodes to generate a candidate cause set. The root cause constraint graph is only used to organize the process logic and define the scope to be verified during this retrieval stage. The inference module does not directly identify the elements in the candidate cause set as the final root cause based on the graph connection relationship of the entity nodes, but instead outputs the candidate cause set to the subsequent data verification process for confirmation.

[0107] Reference Figure 4 For step S30, determining the current effective process path and calculating the dynamic verification window includes the following sub-steps:

[0108] S31. Identify pipeline process connectivity to filter valid process paths. The reasoning module traverses the candidate cause set generated in step S20, and for each candidate cause and its required multi-source evidence, extracts process paths connected to the candidate cause set from the root cause constraint map based on the steady-state equipment opening and closing status before the occurrence of the standard alarm event indicated in the standardized time series data. For the extracted process paths, paths that do not conform to the current operating mode are eliminated to obtain valid process paths. The steady-state equipment opening and closing status before the occurrence of the standard alarm event includes valve opening and closing status, pump operating status, media flow direction, and pipeline isolation status.

[0109] During the screening process, the elimination logic is executed directly, eliminating isolation valves, deactivated pipelines, edges with opposite flow directions to the current medium, and paths corresponding to standby processes that do not belong to the current operating mode, all of which are in a planned shutdown isolation state under the current operating mode. However, the paths of equipment where unplanned state transitions occur at the time of the standard alarm event are retained to ensure that abnormal nodes can be traced. When there are multiple parallel valid process paths between the candidate cause and the alarm trigger point, the inference module calculates the propagation delay of each valid process path and retains the correspondence between each valid process path and its propagation delay, without directly deleting valid paths with larger propagation delays. Subsequently, dynamic verification windows are generated for multi-source evidence that needs to propagate along different valid process paths.

[0110] In cases where the same piece of evidence can reach the target node through multiple valid process paths, the dynamic verification windows corresponding to each valid process path are used together as the valid retrieval time range for that piece of evidence.

[0111] S32. Limited Substitution Estimation and Error Propagation Prevention Constraints under Data Missing Conditions. When calculating propagation delay, real-time parameter data extracted from standardized time-series data is required. This real-time parameter data includes flow rate or velocity data. When the confidence level of the measurement point corresponding to the parameter location is lower than a preset substitution trigger threshold, a limited substitution mechanism is triggered. The preset substitution trigger threshold is calibrated based on sensor baseline accuracy test data and has a configuration range of 0.75 to 0.85. The low confidence marker threshold is used to determine whether the measurement point needs to have its subsequent call priority reduced, while the substitution trigger threshold is used to determine whether to activate the limited substitution mechanism during propagation delay calculation. The low confidence marker threshold and the substitution trigger threshold are independent configuration parameters for different purposes.

[0112] Based on the order of trusted backup points, equivalent points on the same equipment, and constrained associated points on the effective process path, alternative parameter data are retrieved and extracted sequentially to generate alternative estimates and corresponding alternative confidence levels. For trusted backup points or equivalent points measuring the same process location and the same physical quantity, the corresponding effective measurement value can be directly used as the alternative estimate. For constrained associated points on the effective process path, alternative estimates of the target parameter are only allowed to be calculated based on the constraint relationship if the material balance, diversion and convergence, equipment combination, or equivalent conversion relationship between the target point and the associated point is pre-configured in the root cause constraint map, to prevent errors caused by direct substitution without conversion. The alternative confidence level is configured as the confidence level of the measurement point corresponding to the extracted alternative parameter data multiplied by the attenuation coefficient. The attenuation coefficient for trusted backup points or equivalent points ranges from 0.95 to 1.0, and the attenuation coefficient for constrained associated points ranges from 0.75 to 0.9.

[0113] To prevent error divergence, the system executes constraint logic synchronously. Both the attenuation coefficient and the substitution confidence threshold are obtained through regression analysis of historical equipment fault simulation data combined with on-site expert experience, with the substitution confidence threshold configured between 0.5 and 0.6. When the generated substitution confidence level is lower than the substitution confidence threshold, the substitution operation is stopped directly, and the propagation delay on the corresponding relation edge is marked as unavailable. Simultaneously, the system is configured not to use the substitution estimate obtained from a single substitution as the substitution input parameter for other parameter points, thereby preventing continuous error calculation.

[0114] S33. Classified Physical Propagation Delay Calculation and Path Total Delay Deduction. For each relation edge in the effective process path, the inference module performs delay calculations based on the physical constraint parameters bound to the relation edge, including spatial volume and thermal inertia parameters. For evidence characteristics of temperature, composition, and water content transported with the medium, under the condition that the flow rate data is greater than the minimum effective flow rate of the pipeline and the medium flow direction is clear, where the minimum effective flow rate of the pipeline is set by a comprehensive calculation based on the corresponding pipe diameter and the dead zone of the field instrument measurement, fluid delay deduction calculations are performed. The calculation logic is expressed as follows:

[0115] ;

[0116] in, Indicates the delay time corresponding to the relation edge; The effective volume of the pipe segment corresponding to the relation edge is taken from the root cause constraint graph configuration information as a spatial volume parameter; The volumetric flow rate is selected from standardized time-series data or the volumetric flow rate corresponding to alternative estimates obtained based on finite substitution mechanisms.

[0117] For relationships involving storage tanks, filters, and heating furnaces with dynamic response characteristics, due to the buffering effect of the container volume and the influence of the equipment's thermal inertia parameters, and combined with the current operating mode indicated in the standardized time series data, the response time of the corresponding equipment pre-configured in the root cause constraint map is directly extracted as the delay time. The response time of the corresponding equipment is obtained through historical step response test records or equipment factory technical specifications. For relationships representing control commands, interlocking actions, and state feedback, the control feedback response time fixed in the root cause constraint map is used as the delay time.

[0118] After calculating the delay time of all relational edges on the effective process path, the first... The propagation delay of an effective process path is obtained by summing the delay times corresponding to all relation edges contained in that effective process path. .

[0119] S34. Causal Time Direction Calibration and Dynamic Verification Window Generation. The inference module determines the causal time direction and reference event time based on the multi-source evidence associated with the current candidate cause, and sets the value of the causal time direction to include -1, 0, and 1. The specific mapping relationship is as follows: -1 is assigned when representing upstream leading cause evidence; 0 is assigned when representing synchronous phenomena; and 1 is assigned when representing downstream influences, control actions, or feedback signals.

[0120] Subsequently, based on the type of multi-source evidence, the time of the first occurrence of the standard alarm event, the time of the control command issuance, the time of the interlock trigger, the time of the occurrence of the confirmed cause, or the time of the completion of the handling action are selected as reference event times. When using the time of the occurrence of the confirmed cause as the reference event time, the confirmed cause must be other nodes confirmed through an independent chain of evidence or a preceding reasoning process. It is prohibited to use the presumed occurrence time of the candidate cause being verified itself as the time reference.

[0121] By combining the reference event time, the propagation delay of the effective process path, and the causal time direction, a time tolerance is introduced to generate a dynamic verification window. The time tolerance is determined by comprehensively considering the data channel acquisition cycle, process fluctuation deviation, inherent sensor response delay error, and the confidence or surrogate credibility of the direct measurement points corresponding to multi-source evidence.

[0122] The determination of time tolerance adopts a pre-configured hierarchical calculation rule: First, the basic time tolerance is determined based on the sampling period of the corresponding evidence channel, the inherent response delay error of the sensor, and the fluctuation range of the propagation time under historical stable conditions. This tolerance at least covers the maximum value of the above uncertain factors. Then, it is graded and corrected according to the confidence or credibility of the corresponding evidence. The lower the credibility, the wider the verification window is increased within the preset maximum range. When a high credibility level is reached, no additional expansion is performed.

[0123] The basic time tolerance, confidence level, and maximum allowable time tolerance for each type of evidence are pre-stored as configuration parameters for the root cause constraint map. When the same multi-source evidence corresponds to multiple valid process paths, sub-validation windows are generated using the propagation delay of each valid process path, and the union of the resulting sub-validation windows is used as the final dynamic validation time range for that evidence. The calculation logic of the dynamic validation window is expressed as follows:

[0124] ;

[0125] ;

[0126] in, Indicates based on the first The sub-validation window obtained from the calculation of each effective process path; Indicates the first The first candidate reason The final dynamic verification window corresponding to multiple sources of evidence is represented by a set of closed intervals containing a lower time limit and a higher time limit, or their union. This represents the total number of corresponding valid process paths; Indicates the first Reference event time corresponding to multiple sources of evidence; Indicates the first The first candidate reason The causal time direction corresponding to multiple sources of evidence takes the value of a constant. Indicates the number obtained from the corresponding calculation. The propagation delay time of each effective process path; Indicates the first The first candidate reason Time tolerance corresponding to multiple sources of evidence;

[0127] This represents the propagation delay term obtained by combining causal time direction correction; This represents the theoretical arrival time reference point obtained after propagation delay correction following the occurrence of the baseline reference event. The generated dynamic verification window provides time boundary constraints for subsequent extraction of time series data segments on the time axis and for evidence feature matching.

[0128] The dynamic evidence retrieval, evidence readiness determination, and comprehensive evidence calculation in step S40 include the following sub-steps:

[0129] S41. Time-series data extraction and feature extraction based on dynamic verification windows. The inference module performs a retrieval query operation on the standardized time-series data generated by the processing module based on the dynamic verification window calculated in step S30. For the data source channel pointed to by multi-source evidence, the time-series data segment whose timestamp falls within the corresponding dynamic verification window, i.e., between the lower and upper time limits of the corresponding dynamic verification window, is extracted. After obtaining the time-series data segment, statistical feature extraction is performed according to the signal type of the multi-source evidence.

[0130] For temperature, pressure, and flow signals with continuous analog attributes, the maximum, minimum, average, and first-order differential rate of change within the time-series data segment are calculated and combined into an analog feature vector. For valve status and pump operating status feedback signals with discrete digital attributes, steady-state Boolean values ​​and state transition edge sequences within the time-series data segment are extracted and used as digital features. This data extraction and feature extraction process eliminates time deviations caused by asynchronous sampling from different data source channels, unifying the verification and comparison benchmark for operating condition data. Conventional data statistical analysis algorithms are used to extract statistical features such as the maximum, average, and first-order differential rate of change within the time-series data segment.

[0131] S42. Multi-source evidence readiness determination and asynchronous waiting mechanism. To avoid misjudgments due to data gaps caused by downstream physical transmission lag, the inference module performs a readiness determination on multi-source evidence. The current system clock time is compared with the upper time limit of the dynamic verification window. When the current system clock time is less than the upper time limit of the dynamic verification window, the corresponding multi-source evidence is determined to be in an unready state, and the verification task containing feature extraction and feature matching logic is pushed into the system's verification task waiting queue for suspension. This unready state only indicates that the data acquisition process for the corresponding multi-source evidence has not yet ended; it is not output as an independent cause determination state to avoid interfering with subsequent cause three-state determination logic.

[0132] In the verification task waiting queue, time conditions and data reception status are polled and compared at fixed clock cycles. When the complete time-series data segment to be verified has been retrieved from the database and the corresponding timestamp has covered the time limit of the dynamic verification window, the multi-source evidence is determined to be in a ready state; or, when the current system clock time is greater than or equal to the set timeout deadline, the waiting is forcibly terminated. For event-type evidence that only requires the occurrence of a specific event and does not require verification of duration, conditions where a specific event does not occur within the window, or subsequent conflict states, if a target event that meets all the judgment conditions is detected in advance within the dynamic verification window, it is allowed to be judged as ready in advance; for evidence involving duration, steady-state maintenance, no-occurrence conditions, cumulative changes, or evidence that needs to exclude subsequent conflict events, it is necessary to continuously observe until the data corresponding to the time limit of the dynamic verification window arrives before it can be judged as ready.

[0133] Once multi-source evidence is determined to be ready or has reached its timeout deadline, it is removed from the verification task waiting queue and unsuspended, activating the subsequent comprehensive evidence scoring calculation process. The timeout threshold of the verification task waiting queue is adaptively determined based on the dynamic verification window of the corresponding evidence, and the timeout deadline is the upper limit of the dynamic verification window plus a preset fault tolerance period.

[0134] The system can set a basic waiting time of 300 to 600 seconds according to business response requirements. However, when the propagation delay calculated based on the effective process path causes the upper limit of the dynamic verification window time to exceed the basic waiting time, the timeout deadline of the corresponding evidence will be extended to a preset fault tolerance time after the upper limit of the dynamic verification window time. The value range of the preset fault tolerance time is usually configured to be 10 to 60 seconds to accommodate occasional network jitter and data entry delays in industrial network transmission. Only when the data necessary to complete the judgment cannot be obtained after the timeout deadline is exceeded will the corresponding evidence be marked as timeout invalid and removed from the queue to participate in subsequent calculations.

[0135] S43. Multidimensional Feature Comparison and Comprehensive Evidence Score Calculation. For multi-source evidence in a ready state, the inference module extracts the corresponding analog or digital feature vectors and performs multidimensional feature comparison. The extracted feature data is logically verified against the preset constraint attribute data in the root cause constraint map. The constraint attribute data covers the direction of change, the magnitude of change, the duration, the point credibility constraint, the upstream and downstream process relationship, the manual operation record, and the interlocking status constraint. If the feature data meets the requirements of all constraint attribute data, the feature matching status value of the corresponding multi-source evidence is assigned a value of 1; if any constraint condition is not met or the feature cannot be extracted for comparison due to data loss or timeout, the value is assigned a value of 0.

[0136] After feature matching of multi-source evidence is completed, the comprehensive evidence score of the current candidate cause is calculated by combining the confidence scores of the measurement points output by the processing module. Since different multi-source evidences contribute differently to confirming the candidate cause, the system assigns a corresponding weight coefficient to each piece of multi-source evidence. The root cause constraint map reads the configuration data and, for the same candidate cause, divides all positive evidence into a key evidence group and a supporting evidence group. The total weight of the key evidence group is configured to be 0.6 to 0.8, and the total weight of the supporting evidence group is configured to be 0.2 to 0.4, with the sum of the two groups being 1. When the same evidence group includes multiple pieces of evidence, individual weights are assigned within the total weight range of the evidence group based on the importance of each piece of evidence. Simultaneously, for each piece of conflicting evidence, an independent conflict weight coefficient is configured based on its negation effect, with a value ranging from 0.5 to 1.0.

[0137] The weighting coefficients and allocation ranges of the aforementioned pieces of evidence were determined using the analytic hierarchy process (AHP) combined with expert experience scoring, or obtained through regression analysis of historical similar confirmed case data. The calculation logic for the comprehensive evidence score is expressed as follows: (The weighted summation of data parameters extracted from various sources of evidence is then applied.)

[0138] ;

[0139] in, Indicates the current number A comprehensive evidence score for each candidate cause; This represents the normalization function, used to perform interval truncation on the result after subtracting the conflict evidence penalty score from the positive evidence score: when the calculated result is less than 0, the comprehensive evidence score is set to 0; when the calculated result is between 0 and 1, the calculated result remains unchanged; and when the calculated result is greater than 1, the comprehensive evidence score is set to 1. Indicates the current number The total number of supporting and key pieces of evidence associated with each candidate cause; Indicates the current number The total number of conflicting pieces of evidence associated with each candidate cause; Indicates that for the first The candidate reason is set as the first Weighting coefficients corresponding to key pieces of evidence or supporting evidence; Indicates that for the first The candidate reason is set as the first Weighting coefficients corresponding to conflicting pieces of evidence; Indicates the first The candidate cause is associated with the first Feature matching status values ​​of key evidence or supporting evidence;

[0140] Indicates the first The candidate cause is associated with the first For conflict evidence, the conflict intensity value is set to 1 when the observed state meets the pre-set conflict conditions in the root cause constraint map, and 0 when it does not. For conflict conditions with continuous numerical attributes, conflict initiation boundaries and complete conflict boundaries are pre-set for the corresponding conflict evidence in the root cause constraint map. The conflict intensity value is 0 when the observed feature does not reach the conflict initiation boundary, and 1 when it reaches or exceeds the complete conflict boundary. When it is between the conflict initiation boundary and the complete conflict boundary, a linear mapping is performed according to the position ratio of the observed feature relative to the two boundaries, so that the conflict intensity value is between 0 and 1. Indicates the first The credibility of the actual data corresponding to key evidence or supporting evidence is selected. For directly measured points, the confidence level of the measurement point is selected from the corresponding data source channel output by the processing module. For alternative points, the alternative confidence level is selected. Indicates the first The credibility of the actual data corresponding to each conflicting piece of evidence, and the selection logic and... same;

[0141] This indicates the positive score of a single key piece of evidence or supporting evidence after adjustments for credibility and weight. This represents the negative penalty score after adjusting for credibility and weighting of a single piece of conflicting evidence; This represents the sum of all positive scores; This represents the sum of all negative penalty scores.

[0142] For critical evidence that is defined as indispensable in the root cause constraint map, if there is a lack of directly verifiable measurement data, reliable feedback signals, or manual confirmation records, it is prohibited to rely solely on alternative estimates obtained through limited substitution mechanisms to establish the critical evidence.

[0143] For conflict evidence whose conflict intensity value reaches the conflict severity threshold, the conflict severity threshold is calculated based on the confidence interval of historical false alarm conflict data, and is usually configured to be between 0.7 and 0.85. A rejection flag with a Boolean value of true is generated and output separately. This rejection flag with a Boolean value of true is then sent to subsequent steps to perform a direct denial ruling. When the conflict severity threshold is not reached, a rejection flag with a Boolean value of false is output. This flag does not trigger a denial ruling, and the corresponding conflict evidence is only used as a penalty item in the formula to participate in the comprehensive evidence scoring calculation.

[0144] Reference Figure 5 The steps S50, including the determination of the three states of causes, the ranking of root causes, the deduction of the scope of influence, and the assessment of risks, include the following sub-steps:

[0145] S51. Three-state determination of causes based on exclusion flags and comprehensive evidence scores. The inference module receives the comprehensive evidence score, feature matching status, and exclusion flag for each candidate cause output from step S40. After the key evidence corresponding to a candidate cause meets the judgment criteria, the inference module introduces a three-state determination mechanism, ultimately classifying the candidate cause status into three categories: confirmed cause, excluded cause, and unmonitorable cause. For each candidate cause, the system executes multi-branch logic:

[0146] If the required key evidence is available, but there is any of the following situations: key changes have not occurred, the direction of change is opposite to the expectation, the magnitude or duration of change does not meet the requirements, the causal time relationship is not established, the current effective process path does not support the candidate cause, or there is strong conflicting evidence, the reasoning module will mark the corresponding candidate cause as an excluded cause and remove the excluded cause from the candidate cause set to terminate the subsequent inference.

[0147] If all key judgment conditions are met, the comprehensive evidence score is greater than or equal to 0.85, and there is no strongly conflicting evidence, the reasoning module will mark the corresponding candidate cause as a confirmed cause; the 0.85 threshold for determining a confirmed cause is obtained by statistical fitting of the score distribution of historical confirmed cases.

[0148] If the corresponding candidate cause cannot be confirmed or eliminated due to missing necessary parameter points, insufficient confidence of key measuring points and inability to be effectively replaced, missing feedback signals from underlying equipment, or unavailable on-site manual confirmation data, the reasoning module marks it as an unmonitorable cause and records it as a subtype of unavailable necessary monitoring data.

[0149] If there are no conditions that would exclude a candidate cause, but the overall evidence score is still below the threshold for determining a cause after all relevant evidence has been prepared, and the existing evidence is insufficient to form a definitive conclusion, then the corresponding candidate cause will be classified as an unmonitorable cause and recorded as a subtype of insufficient existing valid evidence. Unmonitorable causes include the aforementioned subtypes of unavailable necessary monitoring data and insufficient existing valid evidence.

[0150] In particular, when essential key evidence only has alternative estimates and lacks independent verifiable direct measurement data, reliable equipment feedback, or manual confirmation records, the alternative estimates only serve as supplementary evidence in the scoring and cannot independently meet the verification conditions for determining the cause; if no other independent evidence is available to confirm or exclude the cause, it is classified as an unmonitorable cause.

[0151] S52. Root Cause Hierarchical Ranking Based on Multi-Dimensionality. For a set of candidate causes belonging to definite causes, when there are multiple causes, the inference module performs hierarchical ranking based on the causal topological position of each definite cause in the current effective process path, the time of cause occurrence, the comprehensive evidence score, and the number of accompanying anomalies that can be explained. Among them, causes that can explain other definite causes in terms of causal relationship or are located upstream of other definite causes are given priority as primary root causes. When there is no clear causal upstream and downstream relationship, the ranking priority is determined according to the cause occurrence time, the comprehensive evidence score, and the number of accompanying anomalies that can be explained, and then divided into primary root causes, secondary causes, and accompanying causes.

[0152] S53. Forward extrapolation based on the impact range of effective process paths. The inference module takes the node corresponding to the main root cause as the starting point of the extrapolation, and performs graph traversal forward extension extrapolation along the current effective process path and effective interlocking relationships in the root cause constraint graph.

[0153] During the traversal and inference process, if a closed physically isolated node, an invalid relation edge that does not meet the current operating conditions, a node already visited in the current traversal, or the inference depth reaches the preset maximum propagation range, the inference module stops expanding the target branch. The preset maximum propagation range is set based on the topological complexity and physical isolation boundaries of the station's process units. For predictions of general operational impacts, a preset depth of 3 to 5 process equipment nodes can be used as the standard display or retrieval depth.

[0154] When the inference path involves safety interlocking devices, critical hazardous equipment, or safety-critical relationship edges marked in the root cause constraint graph, the inference continues to the effective physical isolation node, process unit boundary, or pre-configured safety analysis boundary, instead of using the conventional search depth as a mandatory termination condition. Based on the effective traversal range, the inference module generates the inference impact range, which is used to output the predicted affected subsequent equipment, process units, and subsequent derived alarms to the decision module. The specific retrieval calculation mechanism of the forward traversal search is implemented using a conventional graph data query algorithm.

[0155] S54. Event Risk Determination. After obtaining the main root cause and the impact range generated in step S53, the reasoning module obtains the initial alarm level, process parameter deviation, abnormal state duration, importance of equipment associated with the main root cause, comprehensive evidence score of the main root cause, and inferred impact range from the standard alarm event and root cause constraint map. The comprehensive evidence score and abnormal state duration are weighted to characterize the certainty score of the event occurrence, and the initial alarm level, process parameter deviation, importance of associated equipment, and inferred impact range are weighted and summed to characterize the severity of the event consequences.

[0156] Subsequently, the inference module normalizes the deterministic score and severity to preset grid intervals, and then uses the normalized deterministic score and severity as two-dimensional coordinates to match with a preset risk assessment matrix, generating an event risk level. The preset risk assessment matrix is ​​divided into multiple assessment grids by the intersection of the deterministic score axis and the severity axis, with each grid pre-configured with a corresponding level label; the event risk level is divided into one of the following: alert level, attention level, important level, urgent level, and critical level. The event risk level is only one of the input elements for subsequent safety constraint verification and does not directly determine the execution authority of on-site automated actions.

[0157] Reference Figure 6 The process of handling security constraints, recovery verification, and controlled knowledge update in step S60 includes the following sub-steps:

[0158] S61. Generation of Disposal Strategies Based on Primary Root Causes and Risk Levels. The decision-making module obtains the primary root causes, scope of impact, and event risk level. It locates the corresponding cause nodes in the root cause constraint map and extracts corresponding disposal suggestions and control instructions based on the disposal rules associated with the cause nodes. The extracted disposal suggestions and control instructions include automated actions involving pump start-up and shutdown, furnace start-up and shutdown, regulating valve switching, process switching, sewage backwashing, and chemical dosage adjustment, as well as manual operation prompts for on-site inspection verification, maintenance dispatch, and emergency response. The decision-making module formats and encapsulates the extracted content into a standardized disposal list and automatically triggers associated inspection task orders, maintenance work orders, or emergency response plan linkage instructions based on the event risk level. At the same time, it adds preventive isolation measures for downstream nodes expected to be affected.

[0159] S62. Safety constraint verification based on multi-dimensional comprehensive factors. Before outputting control commands to the underlying devices, the decision module performs independent safety constraint verification. Verification items include red-line constraint checks for preventing pressure buildup, dry burning, and evacuation, and comprehensively verify user operation permissions, current physical status of the equipment, writable status of control points, available status of feedback points, confidence level of key measurement points, interlock status, process mutual exclusion relationships, safety boundaries, and effective lifespan window of commands.

[0160] Based on the verification results and event risk levels, the decision module outputs one of the following safety constraint verification results: allow execution, require manual confirmation, block execution, or only provide a prompt. When the safety constraint verification result is allow execution, the decision module allows the control command issuance process to begin. When the safety constraint verification result is require manual confirmation, the control command remains in a pending confirmation state. After an operator with the appropriate permissions completes the confirmation, the current physical state of the equipment, the interlock status, and the effective lifespan window of the command are re-verified. Only after the re-verification is passed can the control command issuance process begin. When the safety constraint verification result is block execution or only provide a prompt, no control request that could cause the equipment to move is sent to the industrial control system. In the only prompt state, only handling suggestions and manual operation prompts are output.

[0161] If there is insufficient key evidence, unmonitorable causes, or unreliable key control feedback, the safety constraint verification result will be forcibly downgraded to require manual confirmation or execution blocking. After the verification permission is issued, the underlying control of the field equipment will still be executed by the existing industrial control system according to the original permissions and hardware interlocking logic, without changing the safety interlocking protection settings of the industrial control system. While outputting the safety constraint verification result and issuing corresponding control commands or prompts based on the verification result, the decision module simultaneously generates a structured evidence chain. The structured evidence chain is organized in a sequential logical order including phenomena, evidence, rules, conclusions, and actions, clearly recording the point values ​​involved in the judgment, the confidence level of the measurement points, the time window, the matching spectrum rules, conflicting evidence, and missing monitoring items.

[0162] S63. Recovery Verification and Controlled Knowledge Version Closed Loop Based on Execution Results. After automated or manual on-site handling is completed, the decision module initiates the recovery verification mechanism to continuously track the recovery status of the operating conditions after the control command is issued to confirm whether the alarm phenomenon has been eliminated, whether the key process parameters remain stable within the recovery observation window, whether the physical feedback of the equipment is consistent with the handling action, whether the upstream and downstream operating conditions are coordinated, and whether new associated alarms have been triggered. The recovery observation window is jointly determined based on the sampling period of the industrial control system and the dynamic response time of the equipment and effective process path involved in this handling, and its duration is not less than 10 to 20 times the sampling period of the industrial control system. When the pre-configured response time of the corresponding equipment, the propagation delay of the relevant effective process path, or the stable holding time specified in the recovery standard is greater than the basic observation duration, the longer of these times is used as the basis for determining the recovery observation window to ensure that the equipment and process have sufficient response time before the recovery result judgment is executed.

[0163] Subsequently, the system aggregates manually diagnosed results, inspection records, maintenance work orders, root cause analysis accuracy evaluations, and recovery effect data to generate update candidate data. For judgment items where deviations have occurred, the decision module submits update requests for the graph parameters within the root cause constraint graph to perform controlled knowledge updates. Controlled knowledge updates only affect the evidence weights, cause judgment thresholds, time tolerances, applicable parameters of delay models, and historical reliability of inference rules used in root cause analysis and evidence reasoning; they do not modify safety interlock protection settings, hardware interlock logic, equipment protection settings, user safety permissions, or safety red-line constraints in industrial control systems and safety instrumented systems. The generated update request compilation is a new version of data with update basis and applicable scope, which is released and takes effect after manual review and approval by authorized personnel. The system supports reverting graph parameters to the previous valid version in case of anomalies, forming a traceable alarm handling and knowledge update closed loop.

[0164] Reference Figure 7 and Figure 8 This application example uses a sudden drop in outlet pressure alarm event of an export pump in the crude oil export process of an oilfield station as an example to illustrate the specific execution process of a knowledge graph-based oilfield station alarm root cause analysis and handling system and method. The crude oil export process of an oilfield station includes export storage tanks, filters, main and backup export pumps, heaters, and corresponding export pipelines.

[0165] The data acquisition module obtains multi-source operational data from the industrial control system and related production management, equipment management, or maintenance data sources. The underlying communication protocols used in the industrial control system include Modbus TCP, OPC UA, and Profibus-DP. The acquisition module acquires continuous analog data of the export pump outlet pressure, export pump inlet pressure, and pipeline flow rate via the OPC UA protocol, and discrete state channel data of the export pump inverter operating frequency, export pipeline valve opening feedback, and pump start / stop status via the Modbus TCP protocol. It also simultaneously reads historical fault and maintenance dispatch records from the equipment management system.

[0166] The processing module receives multi-source runtime data and performs time axis alignment, establishing a unified time scale axis with a fixed time interval of 1 second. For missing data nodes caused by the polling mechanism of the underlying device communication protocol, linear interpolation is used to supplement the missing nodes on the unified time scale when the interval between adjacent valid sampling points does not exceed the preset maximum interpolation interval. The processing module performs dead value marking by calculating the numerical variance of each channel within the sliding time window.

[0167] When the external pump is in operation and the variance of its outlet pressure data channel remains zero for more than the set 0.1% range dead zone threshold, this time series data segment is marked as a frozen dead value and is not included in the verification calculation. The processing module outputs standardized time series data and corresponding measurement point confidence levels based on weighted calculations. When the external pump outlet pressure is detected to drop from 1.5 MPa to 0.8 MPa within the observation period, a standard alarm event is generated, including the trigger time and the alarm characteristics of abnormally low external pump outlet pressure. This abnormal drop point is the... Figure 7 The alarm trigger point for the outlet pressure of the external pump is indicated in the diagram.

[0168] The inference module instantiates the entity and node network of the root cause constraint graph using the Neo4j graph database. The inference module extracts alarm trigger nodes from standard alarm events as initial anchor points and executes a breadth-first search algorithm along the upstream and downstream connections of the crude oil export process and the causal relationship edges, with a search depth limit of 4 layers.

[0169] During graph traversal, the system compares operating parameters and skips edges that do not meet the edge validity conditions to narrow the traversal range, ultimately generating a candidate cause set covering upstream tank liquid level being too low, inlet filter blockage, inverter speed reduction due to failure, and downstream pipeline rupture. The storage mechanism of the graph database and the node traversal implementation process of the breadth-first search algorithm can be implemented using conventional graph data processing methods by those skilled in the art; the underlying addressing and graph traversal logic are well-known technologies in the field and will not be elaborated upon here.

[0170] Based on the steady-state equipment on / off status and fluid flow direction prior to the occurrence of the standard alarm event indicated in the standardized time-series data, the inference module eliminates standby export pump branch pipelines in a deactivated / isolated state involved in the candidate cause set, and filters to generate effective process paths corresponding to the main export process. The physical constraint parameters bound to the edges of the effective process paths are extracted to perform propagation delay calculations. Taking the effective process path from the inlet filter to the export pump as an example, the effective volume of the obtained pipe segment is extracted. The volumetric flow rate is 0.5 cubic meters. It is 0.0417 cubic meters per second. (Based on the formula...) The calculated delay time caused by the physical transport of this pipeline segment is 11.99 seconds.

[0171] The total propagation delay is formed by adding the inherent thermal inertia time of the equipment and the delay times of each pipe segment along the path. For the upstream leading cause evidence of inlet filter blockage, the causal time direction... Assign a value of -1. (Refer to the event time.) Substitute into the formula This includes introducing a time tolerance that covers both the inherent response delay of the sensor and the attenuation of the measurement point confidence level. Generate a sub-validation window in the form of a closed interval set to limit the feature search range. and take it as Figure 7 The dynamic verification window is defined in the middle.

[0172] Within the dynamic verification window, the inference module extracts time-series data segments corresponding to the inlet pressure of the external pump and the motor current from the standardized time-series data, mapping them with the relative time axis (seconds) as the horizontal axis and the normalized amplitude as the vertical axis. When the current system time has not yet reached the upper limit of the dynamic verification window, the corresponding matching task is added to the verification task waiting queue. Upon reaching the upper limit, the system awakens the verification process, extracting the first-order differential rate of change of the inlet pressure of the external pump and the peak value of the motor current fluctuation to form a multi-dimensional analog feature vector. This... Figure 7 The corresponding fitted inlet pressure characteristic waveform and motor current characteristic waveform are shown in the figure.

[0173] Feature matching comparison results show that the inlet pressure drop corresponding to inlet filter blockage and the motor current fluctuation pattern satisfy the preset constraint attributes of the spectrum, and the feature matching value is... The value is assigned to 1. The evidence of an excessively low liquid level in the upstream storage tank failed to extract the expected liquid level decline characteristic, thus generating a corresponding conflict intensity value. This value will be converted to Figure 8 The corresponding conflict rejection penalty score. This is combined with the weighting coefficients set for each source of evidence. Based on the credibility of actual data, the reasoning module performs a weighted summation and conflict deduction calculation of the comprehensive evidence score. Its positive matching part forms a positive feature summation score, and finally calculates the normalized comprehensive evidence score.

[0174] The reasoning module determines the cause of action based on the normalized comprehensive evidence score output. (Refer to...) Figure 8 , Figure 8 The chart displays a multi-dimensional comprehensive evidence score comparison of candidate causes. The horizontal axis lists the four candidate causes mentioned above, and the vertical axis represents the evidence score values ​​for each. The candidate cause of low upstream tank liquid level was classified as an excluded cause and directly eliminated due to the presence of a strong exclusion indicator and the corresponding normalized comprehensive score dropping to 0. Inverter malfunction causing speed reduction was also determined to be an excluded cause because the frequency feedback was normal and there were no accompanying electrical abnormalities.

[0175] The comprehensive evidence score for inlet filter blockage reached 0.88, exceeding the 0.85 threshold for determining the cause (obtained by fitting historical data distribution), thus confirming it as a definite cause. The reasoning module used this definite cause as a starting point to extend the deduction forward along the interlocking constraint edges. The deduction results indicated that pump inlet cavitation might further damage the mechanical seal of the external pump. By integrating the affected area, process deviation magnitude, and duration of the anomaly, and mapping it to a pre-set two-dimensional risk assessment matrix, the event risk level was output as "attention level."

[0176] The decision-making module retrieves the set of handling instructions associated with the inlet filter blockage and generates a standardized handling list. The list includes a data packet containing control instructions to stop the main external pump and switch to start the standby filter and pump unit. Before execution, the decision-making module verifies the operator's authorization level, the writable status of the inlet and outlet valves of the standby pipeline section, and performs red-line constraint checks to prevent pressure buildup and cavitation.

[0177] After verification confirms the safety constraint check result is valid and execution is permitted, a structured evidence chain is generated synchronously. The control command is encapsulated into a register write operation message based on the Modbus TCP protocol and sent to the field programmable logic controller to execute the valve group and pump group switching action. After the action is triggered and serves as a recovery observation window for 10 times the sampling period of the industrial control system, the decision module extracts the inlet pressure parameter to perform recovery verification. After the verification is successful, controlled knowledge updates are performed on the local evidence weights that showed deviations in the simulation.

[0178] The technical effects of the present invention will be further discussed in conjunction with the embodiments and the above-mentioned accompanying drawings and data. Figure 7 The system extracts waveform comparison features based on time-series data with dynamic verification windows, and addresses the issue of sudden changes in the outlet pressure of the external pump. This breaks away from the limitations of traditional alarm processing, which simply extracts alarm trigger section parameters for logical verification.

[0179] Based on physical constraint parameters such as spatial volume attached to the root cause constraint map, the inherent hysteresis time of fluid conduction within the pipe section is calculated and then compensated in reverse to form... Figure 7 The dynamic verification window interval within this window shows non-stationary changes in the inlet pressure characteristic waveform and the motor current characteristic waveform, confirming the mechanism's ability to isolate the medium conduction hysteresis error and lock in the true leading physical laws.

[0180] according to Figure 8 The multi-dimensional comprehensive evidence score comparison of the candidate causes presented shows that relying solely on the accumulation of positive evidence can easily lead to misdiagnosis due to homologous anomalies. The candidate cause of low liquid level in the upstream storage tank obtained a positive feature accumulation score of 0.25 due to process fluctuations, but the conflict condition retrieval introduced by the system extracted a strong conflict rejection penalty score of 0.85. After calculation and truncation, the final normalized comprehensive evidence score was reduced to zero, thus preventing the extension of erroneous decision branches.

[0181] In contrast, inlet filter blockage exhibited a high positive matching score without any penalty for conflicting evidence, ultimately achieving a determination score exceeding the preset threshold of 0.85 for identifying the cause. Other causes, such as downstream pipeline rupture, were also effectively filtered out based on objective scores. This scoring mechanism, combined with the anti-proliferation isolation attenuation rule for low-confidence data, establishes a highly robust diagnostic benchmark for addressing industrial sensor degradation and sporadic interference.

[0182] After identifying the core cause of the system state deviation, the decision-making and execution phase is no longer subject to hard-lock intervention in the underlying system's security configuration. Instead, a security constraint verification strategy based on multi-dimensional factors is adopted. Before sending control request packets, the system completes mutual exclusion verification and evaluation of physical red-line constraints and manually authorized control point status. The deductive decision-making logic is left to the station management control layer, while direct execution of protection actions relies on the original protection settings of the basic equipment. This effectively closes the abnormal state handling process without weakening the deterministic mechanism of the basic safety interlock level response, constructing a high-safety-margin automated auxiliary production closed-loop control framework.

[0183] While specific embodiments of the present invention have been described above, those skilled in the art should understand that these specific embodiments are merely illustrative. Those skilled in the art can omit, substitute, and modify the details of the above methods and systems in various ways without departing from the principles and essence of the present invention. For example, combining the above method steps to perform substantially the same function and achieve substantially the same result using substantially the same method falls within the scope of the present invention.

Claims

1. A method for root cause analysis and handling of oilfield station alarms based on knowledge graphs, characterized in that, Includes the following steps: Perform standardization processing on multi-source operational data to output standardized time-series data, calculate the measurement point confidence of the standardized time-series data, and generate standard alarm events containing alarm features; A root cause constraint map is constructed based on process flow and equipment ledger information, and a candidate cause set is generated by retrieving the root cause constraint map using the alarm features; Valid process paths are selected from the candidate cause set, and the propagation delay of the valid process paths is calculated based on the confidence level of the measurement points. Acquire multi-source evidence, determine the causal time direction and reference event time of the multi-source evidence, and generate a dynamic verification window for the multi-source evidence based on the causal time direction, the propagation delay of the effective process path, and the reference event time. The execution features of the time-series data segments corresponding to the multi-source evidence are extracted within the dynamic verification window for verification, and a comprehensive evidence score is calculated. Based on the comprehensive evidence score and the pre-set conflict and exclusion rules in the root cause constraint map, the candidate cause set is classified into definite causes, and the scope of influence is deduced along the root cause constraint map to determine the event risk level; Based on the event risk level, a handling suggestion and control instruction are generated, the control instruction is issued to execute the handling, recovery verification is performed, and the graph parameters in the root cause constraint graph are updated with controlled knowledge.

2. The method for root cause analysis and handling of oilfield station alarms based on knowledge graphs according to claim 1, characterized in that, Perform standardization processing on multi-source operational data to output standardized time-series data, calculate the measurement point confidence of the standardized time-series data, and generate standard alarm events containing alarm characteristics, including: Acquire the multi-source operational data and establish a unified time scale axis for the sampling sequences of different data source channels in the multi-source operational data; Perform a dimensional normalization operation to convert the engineering units extracted from different data source channels into a unified standard physical unit; Perform a dead value marking operation to mark the sampled sequence as a frozen dead value, and complete the standardization process to obtain the standardized time series data; Extract the quality code and time series stability from the standardized time series data, and calculate the measurement point confidence level corresponding to the data source channel based on the extracted quality code and time series stability. Based on the standardized time-series data, the monitoring conditions generate the standard alarm event, which includes the trigger time and the alarm characteristics.

3. The knowledge graph-based root cause analysis and handling method for oilfield station alarms according to claim 1, characterized in that, A root cause constraint map is constructed based on process flow and equipment ledger information. The alarm features are then used to retrieve and generate a candidate cause set from the root cause constraint map, including: Based on the process flow and equipment ledger information, construct the node set of station equipment and pipeline processes required for the root cause constraint graph, and establish relationship edges between the node sets to associate different entities; Write the edge validity condition for the relationship edge; Extract the alarm features contained in the standard alarm event, and map the triggering device and triggering point in the alarm features as initial anchor points to the entity nodes corresponding to the root cause constraint map; Using the initial anchor point as the starting point, the search logic is executed by comparing the edge validity conditions of the relation edge with the current running mode of the standard alarm event record. Relationship edges that do not meet the edge validity conditions are skipped, and the candidate cause nodes mapped to the end point of the search path are collected to generate the candidate cause set.

4. The knowledge graph-based root cause analysis and handling method for oilfield station alarms according to claim 1, characterized in that, The selection of effective process paths for the candidate cause set includes: Based on the steady-state equipment on / off status before the occurrence of the standard alarm event indicated in the standardized time series data, process paths connected to the candidate cause set are extracted from the root cause constraint map. For the extracted process paths, paths that do not conform to the current operating mode are discarded to obtain the valid process paths.

5. The method for root cause analysis and handling of oilfield station alarms based on knowledge graphs according to claim 1, characterized in that, The propagation delay of the effective process path calculated based on the confidence level of the measurement points includes: Extract the real-time parameter data corresponding to the effective process path from the standardized time series data; When the confidence level of the measurement point corresponding to the real-time parameter data is lower than the preset substitution trigger threshold, the associated substitution parameter data is retrieved from the root cause constraint map to generate a substitution estimate. The real-time parameter data or the alternative estimated value is selected based on the confidence level of the measurement point. For each relation edge in the effective process path, the corresponding delay time is determined based on the selected real-time parameter data or the alternative estimated value. The delay times corresponding to all relation edges included in the effective process path are numerically summed to obtain the propagation delay of the effective process path.

6. The method for root cause analysis and handling of oilfield station alarms based on knowledge graphs according to claim 1, characterized in that, Acquire multi-source evidence, determine the causal time direction and reference event time of the multi-source evidence, and generate a dynamic verification window for the multi-source evidence based on the causal time direction, the propagation delay of the effective process path, and the reference event time, including: Based on the causal time direction, the propagation delay of the effective process path, and the reference event time, and by introducing time tolerance, the lower time limit and the upper time limit of the dynamic verification window are determined. The dynamic verification window corresponding to the multi-source evidence is generated based on the lower time limit and the upper time limit.

7. The method for root cause analysis and handling of oilfield station alarms based on knowledge graphs according to claim 1, characterized in that, Within the dynamic verification window, the execution features of the time-series data segments corresponding to the multi-source evidence are extracted for verification, and a comprehensive evidence score is calculated, including: Extract the time-series data segment corresponding to the multi-source evidence whose timestamp falls within the dynamic verification window; The feature data of the time series data segment is extracted and logically verified with the preset constraint attribute data in the root cause constraint map. Based on the verification results, the feature matching state value and conflict intensity value corresponding to the multi-source evidence are determined. The multi-source evidence is divided into positive evidence and conflicting evidence. A weight coefficient is assigned to the positive evidence, and an independent conflict weight coefficient is configured for the conflicting evidence. A positive score is calculated based on the weight coefficient, feature matching state value, and corresponding test point confidence of a single positive piece of evidence; a negative penalty score is calculated based on the conflict weight coefficient, conflict intensity value, and corresponding test point confidence of a single conflict piece of evidence. The comprehensive evidence score is calculated based on all the positive scores and all the negative penalty scores.

8. The method for root cause analysis and handling of oilfield station alarms based on knowledge graphs according to claim 1, characterized in that, Based on the comprehensive evidence score and the pre-defined conflict and exclusion rules in the root cause constraint map, the candidate cause set is classified to determine the causes, including: The comprehensive evidence score is compared with a preset threshold for determining the cause. If the comprehensive evidence score meets the determination threshold for determining the cause, and the conflict exclusion rule determines that there is no strongly conflicting evidence, then the corresponding candidate cause in the candidate cause set is marked as the determined cause. If the strong conflict evidence is determined to exist according to the conflict exclusion rule, the corresponding candidate cause is marked as an excluded cause and removed from the candidate cause set; If the comprehensive evidence score does not meet the determination threshold for cause and there is no strong conflicting evidence, the corresponding candidate cause will be marked as an unmonitorable cause.

9. The knowledge graph-based root cause analysis and handling method for oilfield station alarms according to claim 1, characterized in that, Extrapolating the scope of influence and determining the event risk level along the root cause constraint map includes: The primary root cause is determined by hierarchical ranking based on the causal topological location of the identified cause in the effective process path, the time of cause occurrence, the comprehensive evidence score, and the number of accompanying anomalies that can be explained. Taking the nodes corresponding to the main root causes as the starting point of the deduction, a graph traversal expansion deduction is performed in the root cause constraint graph, and the deduction influence range is generated based on the effective traversal range. The initial alarm level, process parameter deviation, duration of abnormal state, and importance of the main root cause-related equipment are obtained from the standard alarm events and the root cause constraint map. A deterministic score is calculated based on the comprehensive evidence score and the duration of the abnormal state, and the severity is calculated based on the initial alarm level, the degree of deviation of the process parameters, the importance of the main root cause-related equipment, and the inferred impact range. The event risk level is generated by matching the deterministic score with the severity input to a preset risk assessment matrix.

10. The method for root cause analysis and handling of oilfield station alarms based on knowledge graphs according to claim 1, characterized in that, Based on the event risk level, generate handling suggestions and control instructions, issue the control instructions to execute the handling, perform recovery verification, and update the graph parameters in the root cause constraint graph with controlled knowledge, including: Receive the event risk level and generate the handling suggestion and the control instruction, and perform a safety constraint verification on the control instruction based on the root cause constraint map; When the safety constraint verification result indicates that execution is permitted, the control command is sent to the industrial control system; The recovery verification is performed by tracking the recovery status of the operating conditions after the control command is issued. Based on the recovery effect data obtained from the recovery verification and the accuracy evaluation of the root cause determination, the controlled knowledge update is performed on the graph parameters within the root cause constraint graph.

Images