A process equipment failure prediction and health management method and system

Abstract

The present application relates to the technical field of compressor unit intelligent operation and maintenance, more specifically, the present application relates to a process equipment fault prediction and health management method and system. The method generates health index, fault risk score, risk state label and suggestion work order through multi-source data acquisition, working condition anchor point mapping, equipment group mechanism reference spectrum construction, dynamic effective signal subset screening, equivalent residual package and health factor coupling analysis, causal precursor transfer diagram identification, degradation causal chain reconstruction under missing mode and intervention feedback self-correction. The scheme can reduce cross-condition misjudgment, improve prediction continuity and key path identification reliability under missing interruption scenario.

Description

Technical Field

[0001] This invention relates to the field of intelligent operation and maintenance technology for compressor units, and more specifically, to a method and system for predicting process equipment failures and managing health. Background Technology

[0002] In the oil and gas industry, compressor units are typically key equipment for continuous operation of pressurization, transportation, recovery, and related processes. Their operating status directly affects transportation efficiency, energy consumption, equipment stability, and station safety. Existing compressor unit condition monitoring or fault early warning schemes usually focus on identifying single equipment, single parameters, or local anomalies, making it difficult to simultaneously address issues such as differences among equipment groups, operating condition switching, control action interference, maintenance and recovery impacts, and fluctuations in the quality of multi-source data.

[0003] Existing technologies typically employ a uniform fixed threshold or isolated modeling of single equipment for early warning judgments of compressor units. This can easily lead to cross-equipment and cross-operating condition misjudgments when different equipment have differences in installation conditions, wear levels, media characteristics, operating history, and calibration status. At the same time, due to the lack of unified operating condition constraints for start-up and shutdown phases, load ranges, anti-surge action status, and maintenance and recovery phases, data from different operating phases can easily be directly mixed and compared, resulting in distorted anomaly judgments.

[0004] Furthermore, existing technologies are generally not well-suited to handling data loss and disruption. Monitoring data from compressor units in oil and gas stations comes from complex sources and is frequently affected by sensor anomalies, communication fluctuations, interface delays, and maintenance operations, often resulting in localized data loss, continuous segment loss, missing key variables, or simultaneous multi-source data loss. The lack of a unified mechanism for identifying and reconstructing missing patterns often leads to interruptions in risk assessments, incomparable results, and even undermines the foundation for subsequent anomaly propagation analysis.

[0005] Therefore, we propose a method and system for predicting and managing process equipment failures to address the above problems. Summary of the Invention

[0006] In order to overcome the above-mentioned defects of the prior art, embodiments of the present invention provide a method and system for process equipment failure prediction and health management to solve the problems mentioned in the background art.

[0007] To achieve the above objectives, the present invention provides the following technical solution: a method for process equipment failure prediction and health management, comprising: According to the preset acquisition cycle, the operating data of the target compressor unit is acquired, and control command data, alarm event data, maintenance records, environmental data and historical fault sample data are acquired simultaneously. The operating data includes at least inlet pressure, outlet pressure, inlet temperature, exhaust temperature, flow rate, vibration, speed, current or power, anti-surge valve opening, lubrication-related parameters and cooling-related parameters. Based on the time synchronization tolerance window, start-stop state switching time, load interval switching time, anti-surge action time, and maintenance action time, the operation data, control command data, alarm event data, and maintenance records are time-aligned and operating condition anchor point mapped to generate a unified operating condition anchor point reference axis and a unified event record sequence arranged thereafter. In the unified event record sequence, stable operating segments are identified, and a mechanism reference spectrum is constructed based on the historical effective operating data of the equipment group under the same operating condition anchor point. At the same time, the single-unit correction vector is determined based on the historical offset of the target compressor unit relative to the equipment group, and the causal feasible region between the preceding and following nodes that satisfies the physical propagation order of the compressor unit is defined by the mechanism reference spectrum. Based on the missing status, continuity status, and correlation closure status of each signal within the current monitoring period, a subset of dynamically valid signals is selected from the operational data, and the missing pattern corresponding to the current monitoring period is identified. The equivalent residual packet is calculated based on the dynamic effective signal subset, mechanism reference spectrum and single-machine correction vector. Health factors are extracted based on short time windows and long time windows. The residual-health coupling state matrix is ​​constructed based on the joint change of the equivalent residual packet and health factors. When the residual-health coupling state matrix undergoes a cross-boundary change relative to the reference boundary of the corresponding stable segment and remains monotonically evolving within a long time window, a degenerate segment is generated, and the segment confidence score, coupling state identifier, and precursor feature vector are determined. An initial precursor transition graph is constructed based on the temporal adjacency relationship between degraded segments, the continuity relationship of working condition anchor points, the consistency relationship of residual direction, the maintenance of reversibility, the multi-source support relationship, and the causal feasible region. For candidate edges in the initial precursor transition graph, causal direction identification, confusion factor detection and path pruning are performed under the same working condition anchor point to generate a causal precursor transition graph, and the causal strength coefficient and causal edge confidence are determined for the retained causal edges. When the dynamic effective signal subset does not meet the normal prediction calculation conditions, a degraded causal chain reconstruction is performed based on the missing mode to generate a degraded causal precursor transition map. Based on the causal precursor transition map or the downgraded causal precursor transition map, the dominant causal transition sequence, key causal propagation paths and optimal blocking point solutions are extracted, and a health index, fault risk score and risk status label are generated. Based on the optimal blocking point solution, the recovery benefit estimate and the suggested execution window, a suggested work order is generated. After the recommended work order is executed, intervention feedback data is collected, and self-correction updates are performed on the causal edge credibility, causal strength coefficient, fault rule template and optimal blocking point scheme scoring parameters based on the intervention feedback data, so as to output health management results that meet the risk continuity constraint and physical consistency constraint.

[0008] In a preferred embodiment, the generation of the unified operating condition anchor point reference axis includes: using a standard time series formed by a preset acquisition cycle as a basic index, resampling the continuous sampling data to the basic index, mapping alarm event data and maintenance records to the corresponding index segments according to the trigger time, release time or duration interval, and setting the compressor unit's start-up and shutdown phase, load interval, anti-surge action status and maintenance recovery phase as different types of operating condition anchor points respectively. The mechanism reference spectrum is constructed under the same operating condition anchor point and in the stable operating range, and includes at least the correlation constraint relationship between pressure ratio and flow rate, the correlation constraint relationship between speed and power, the correlation constraint relationship between anti-surge valve opening and flow fluctuation, and the correlation constraint relationship between exhaust temperature and load. The causal feasible region is limited by the mechanism reference spectrum to establish candidate causal edges only when the physical propagation sequence of the compressor unit, the variable coupling closure relationship, and the consistency of the operating condition anchor point are satisfied.

[0009] In a preferred embodiment, the dynamic effective signal subset is determined by calculating the field integrity index, time continuity index, and correlation closure index of each operating signal in the current monitoring period, and including operating signals whose field integrity index is not lower than a first threshold, whose time continuity index is not lower than a second threshold, and which satisfy the correlation closure condition with at least two core mechanism variables into the dynamic effective signal subset. Among them, the core mechanism variables include at least pressure ratio, flow rate, rotational speed, power, and anti-surge valve opening; The missing pattern is obtained by jointly judging the distribution of missing locations, the duration of missing data, the importance level of missing variables, and the number of missing sources, so that normal prediction calculation and downgraded causal chain reconstruction can be carried out in the same residual expression space.

[0010] In a preferred embodiment, the equivalent residual package includes at least pressure ratio residual, flow rate residual, power residual, exhaust temperature residual, vibration deviation, and steady-state recovery residual after control execution; The health factors include at least alarm co-occurrence density, control response hysteresis, start-stop disturbance degree, maintenance recovery rate, cross-source consistency score, and missing continuity index. The residual-health coupled state matrix is ​​obtained by cross-combining the residual terms in the equivalent residual package and the factor terms in the health factor according to the working condition anchor point and time window. The short time window is used to characterize rapid abnormal fluctuations and control response lag, while the long time window is used to characterize the continuous degradation trend and the decay trend of maintenance and recovery effect. The conditions for generating the degraded segment include: the residual-health coupling state matrix undergoes a cross-boundary change relative to the reference boundary of the corresponding stable segment of the operating condition, and within a long time window, it satisfies the following: consistent evolution direction, duration exceeding a preset duration, and no conflict with the causal feasible region.

[0011] In a preferred embodiment, the causal direction identification includes: for any two nodes A and B with candidate edge relationships in the initial precursor transfer graph, constructing a time series causal test model based on the historical values ​​of node A and node B under the same working condition anchor point, to determine whether the historical values ​​of node A have a significant contribution to the prediction result of node B, and calculating the transfer entropy from node A to node B and the transfer entropy from node B to node A respectively. When the transfer entropy from node A to node B is greater than the transfer entropy from node B to node A and the difference between the two exceeds a preset threshold, and the time series causal test model supports that node A has a significant impact on node B, the direction from node A to node B is determined to be the dominant causal direction and the corresponding candidate edge is retained. The confusion factor detection selects at least one or more of the following as potential confusion factor candidates: operating condition anchor point type, ambient temperature, load level, and cumulative equipment operating time. Under the given potential confusion factor candidates, conditional independence tests are performed on the preceding and following nodes corresponding to the candidate edges.

[0012] Causal direction identification is preferably performed under the same working condition anchor point. For any two nodes A and B with candidate edge relationships, multiple candidate lag orders are first selected within a preset time lag range to construct time series causal test models for nodes A and B respectively. The lag order corresponding to the optimal information criterion or the minimum prediction error is taken as the target lag order for the current node pair. Then, the transfer entropy from node A to node B and from node B to node A is calculated under this target lag order. For continuous variables, it is preferable to first discretize them according to a preset number of bins or estimate the probability distribution using kernel density. When the causal test result from node A to node B meets the preset significance level, and the transfer entropy from node A to node B is greater than the transfer entropy from node B to node A and exceeds a preset difference threshold, the candidate edge from A to B is retained; otherwise, the candidate edge is deleted or marked as a weak causal candidate edge.

[0013] In a preferred embodiment, for the candidate edges retained after the conditional independence test, an initial causal edge credibility is generated based on the causal direction identification result, the direction advantage of the transfer entropy, the consistency of the working condition anchor point, the degree of multi-source support and the constraint compliance of the mechanism reference spectrum, and a weighted average is applied to each retained edge in combination with the causal strength coefficient. The dominant causal transition sequence is jointly extracted based on the cumulative path weight, causal strength coefficient, causal edge credibility, and sequence similarity with historical fault sample data under the same working condition anchor point. The optimal blocking point scheme is determined comprehensively based on the risk reduction, implementation cost, implementation difficulty, and estimated recovery benefit of each node or edge in the key causal propagation path.

[0014] In a preferred embodiment, the downgraded causal chain reconstruction includes: when the missing mode is random missing, while keeping the topology of the cause-effect precursor transition graph unchanged, performing local residual compensation on the missing location based on adjacent valid sampling points and adjacent valid operating condition anchor points; When the missing pattern is a continuous segment missing, a bridging causal edge is constructed based on the stable operating segment before and after the missing segment, the coupling state identifier before and after the missing segment, and the dominant causal transition sequence of the previous effective monitoring cycle. When the missing mode is the missing key variable, a substitute variable that satisfies the mechanistic closure relationship with the missing variable is selected, and an equivalent bridging residual is generated by combining the mechanistic reference spectrum and the single-machine correction vector. The key causal chain is then reconstructed based on the equivalent bridging residual. When the missing mode is synchronous multi-source missing, a conservative update is performed based on the remaining dynamic valid signal subset, the dominant causal transfer sequence of the previous valid monitoring period, and the preset risk change boundary, and the recommended execution window is shortened.

[0015] The reconstruction of downgraded causal chains preferably follows the principle of "prioritizing topology preservation, then rebuilding key edges, and finally performing conservative updates." For scenarios where key variables are missing, substitute variables are preferably selected from those that satisfy the mechanistic closure relationship with the missing variable, have stable correlation constraints under the same operating condition anchor point, and meet the requirements for current field completeness and temporal continuity. When multiple substitute variables exist, the variable with the highest historical correlation to the missing variable and the highest cross-source consistency score within the current monitoring period is preferred as the primary substitute variable. For scenarios with continuous missing segments, bridging causal edges are preferably established jointly based on the coupling state identifiers before and after the missing segment, the consistency of residual directions before and after, the continuity relationship of operating condition anchor points before and after, and the dominant causal transfer sequence of the previous effective monitoring period. For synchronous multi-source missing scenarios, conservative updates preferably limit the change in the current fault risk score relative to the risk score of the previous effective monitoring period to no more than a preset risk change boundary, and simultaneously shorten the suggested execution window to avoid outputting overly aggressive conclusions under conditions of large-scale missing data.

[0016] In a preferred embodiment, the self-correction update of the credibility of the causal edge includes: for each causal edge on the key causal propagation path, respectively, the risk reduction amount, alarm expiration time, parameter regression degree, recovery time and actual intervention cost after the corresponding intervention measures are implemented; When the actual risk reduction after the intervention measures corresponding to a certain causal edge are implemented are greater than the preset expected value and the parameter regression degree meets the preset conditions, the causal edge credibility of that causal edge and the score of the corresponding node or edge in the optimal blocking point scheme are increased. When the actual risk reduction after the intervention measures corresponding to a certain causal edge are implemented are lower than the preset expected value or the recovery time exceeds the preset time, the causal edge credibility of that causal edge is reduced, and the weight of the corresponding fault rule template and the score of the optimal blocking point scheme are lowered.

[0017] The intervention feedback self-correction mechanism is preferably triggered after a single suggested work order is completed and the preset observation end conditions are met, or in batches after the cumulative number of intervention feedback samples reaches a preset number. To avoid drastic parameter fluctuations due to a single accidental intervention result, an incremental update method is preferred to correct the causal edge confidence, causal strength coefficient, fault rule template weight, and optimal blocking point scheme scoring parameters, with an upper limit on the single update magnitude. When the number of intervention feedback samples is insufficient, key fields are missing in the feedback results, or the current working condition anchor point differs too much from the historical main anchor point, the self-correction update is preferably suspended, and only the feedback data is retained for subsequent cumulative updates.

[0018] A process equipment failure prediction and health management system, comprising: The data access module is used to acquire the operating data of the target compressor unit equipment group in the oil and gas business according to the preset acquisition cycle, and simultaneously acquire control command data, alarm event data, maintenance records, environmental data and historical fault sample data. The working condition anchor point mapping module is used to generate a unified working condition anchor point reference axis and output a unified event record sequence. The mechanism reference construction module is used to construct the equipment group-level mechanism reference spectrum, generate the single-unit correction vector of the target compressor unit, and define the causal feasible region under the conditions of the same operating condition anchor point and in the stable operating condition segment. The dynamic valid signal filtering and missing pattern identification module is used to filter a subset of dynamic valid signals and identify missing patterns based on field completeness, temporal continuity and correlation closure. The equivalent residual packet generation module is used to generate equivalent residual packets based on the dynamic effective signal subset, the mechanism reference spectrum, and the single-machine correction vector. The coupling state analysis module is used to generate a residual-health coupling state matrix based on the equivalent residual package and health factors and to identify degenerate segments. The initial precursor transition graph construction module is used to construct the initial precursor transition graph based on the temporal adjacency relationship between degenerate segments, the continuity relationship of operating condition anchor points, the consistency relationship of residual direction, the maintenance of reversibility relationship, the multi-source support relationship, and the causal feasible region. The causal pruning and critical path identification module is used to perform causal direction identification, confusion factor detection, conditional independence test, causal strength coefficient calculation, causal edge confidence determination, path pruning, and extraction of dominant causal transfer sequence and optimal blocking point scheme under the same working condition anchor point. The downgraded causal chain reconstruction module is used to perform downgraded causal chain reconstruction based on the missing mode and generate a downgraded causal precursor transition map when the dynamic effective signal subset does not meet the normal prediction calculation conditions. The risk fusion assessment module is used to generate health index, fault risk score, risk status label and suggested work order; The intervention feedback self-correction module is used to update the causal edge confidence, causal strength coefficient, fault rule template weight, and optimal blocking point scheme score based on intervention feedback data after the suggested work order is executed. The system adopts an online hybrid deployment method and includes at least a field edge processing unit and a central analysis and processing unit. The field edge processing unit is used to perform data access, dynamic effective signal filtering, missing pattern identification, downgraded causal chain reconstruction, and conservative update calculation. The central analysis and processing unit is used to perform device group-level mechanism reference spectrum update, causal pruning and critical path identification, intervention feedback self-correction, and risk fusion assessment.

[0019] The technical effects and advantages of this invention are as follows: 1. This invention generates a unified operating condition anchor point reference axis and a unified event record sequence by aligning the execution time of operating data, control command data, alarm event data, and maintenance records, and mapping operating condition anchor points. This enables data from the start-up and shutdown phases, load ranges, anti-surge action states, and maintenance recovery phases to be compared in a consistent context, thereby improving the accuracy of state determination.

[0020] 2. Under the condition of the same anchor point and in the stable operating range, the present invention constructs a mechanism reference spectrum based on the historical effective operating data of the equipment group, and determines the single-unit correction vector by combining the historical offset of the target compressor unit relative to the equipment group, thereby taking into account both the common laws of the equipment group and the individual differences of the target compressor unit, and reducing the risk of misjudgment.

[0021] 3. This invention filters a subset of dynamic and effective signals by using field integrity indicators, time continuity indicators, and correlation closure indicators. It also identifies missing patterns by jointly discriminating the distribution of missing locations, the duration of missing data, the importance level of missing variables, and the number of missing sources. This provides a unified residual expression basis for normal prediction calculations and downgraded causal chain reconstruction, thereby improving the continuity of results.

[0022] 4. This invention calculates the equivalent residual package, extracts health factors under short and long time windows, and constructs a residual-health coupled state matrix, thereby simultaneously reflecting rapid fluctuations, control response lag, continuous degradation, and maintenance recovery decay, improving the ability to identify progressive degradation and compound anomalies.

[0023] 5. This invention constructs an initial precursor transition map based on degenerate fragments, and further performs causal direction identification, confounding factor detection, and path pruning, thereby reducing spurious correlation paths and improving the reliability of identifying dominant causal transition sequences and key causal propagation paths.

[0024] 6. This invention determines the optimal blocking point scheme based on the risk reduction, implementation cost, implementation difficulty, and recovery benefit estimate corresponding to each node or edge in the key causal propagation path, and generates suggested work orders and suggested execution windows, thereby improving the pertinence of operation and maintenance decisions and the feasibility of engineering.

[0025] 7. This invention collects intervention feedback data after the suggested work order is executed, and performs self-correction updates on the causal edge credibility, causal strength coefficient, fault rule template and optimal blocking point scheme scoring parameters, thereby enabling the system to have continuous learning and continuous optimization capabilities, and improving the reliability of subsequent health management results. Attached Figure Description

[0026] Figure 1 This is a system framework diagram of the present invention. Detailed Implementation

[0027] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0028] Reference Figure 1 A method for predicting and managing the health of process equipment failures, comprising: The system acquires operating data of the target compressor unit according to a preset acquisition cycle, and simultaneously acquires control command data, alarm event data, maintenance records, environmental data, and historical fault sample data. Data acquisition is accomplished using a combined approach of "field acquisition terminal + control system interface + business system retrieval + historical database backfilling," thereby ensuring that data from different sources can be used uniformly within the same monitoring framework.

[0029] Operational data is directly collected by field sensors deployed on the target compressor unit itself and its auxiliary pipelines, lubrication units, and cooling units.

[0030] The inlet and outlet pressures can be obtained by pressure transmitters installed on the inlet and outlet pipelines; the inlet and outlet temperatures can be obtained by thermocouples, resistance temperature detectors (RTDs), or temperature transmitters; and the flow rate can be obtained by orifice flow meters, vortex flow meters, ultrasonic flow meters, or mass flow meters. Vibrations can be detected by vibration sensors installed in bearing housings, housings, or critical support locations; Rotational speed can be obtained from a rotational speed probe, encoder, or speed measuring device; Current or power can be obtained from the motor-side current transformer, power transmitter, or power distribution monitoring unit; The opening degree of the anti-surge valve can be obtained from the valve positioner, the actuator feedback module, or the valve position feedback signal in the control system; Lubrication-related parameters can be obtained from oil pressure sensors, oil temperature sensors, oil level switches, and oil quality monitoring units; Cooling-related parameters can be acquired by cooling water inlet and outlet temperature sensors, cooling medium flow meters, and cooling pressure sensors. These field sensors are connected via PLC, DCS, SCADA, RTU, or edge acquisition gateways and uploaded to the monitoring system via industrial communication protocols.

[0031] Control command data is read directly from the compressor unit's control system, including start-up commands, stop commands, loading / unloading commands, speed adjustment commands, anti-surge control commands, valve opening / closing commands, bypass switching commands, etc. This type of data typically does not require additional sensors; it is output by the DCS, PLC, or configuration monitoring system via a communication interface, and then captured by edge nodes or central analysis nodes according to timestamps. To ensure the correlation between control commands and equipment responses, corresponding execution feedback signals are also acquired and stored along with the operating data.

[0032] For alarm event data, it is retrieved from the alarm management system, DCS alarm database, SCADA event database, or unit protection system. The retrieved information includes at least the alarm number, alarm type, trigger time, release time, alarm duration, and alarm level. In practice, new alarms can be received in real time via event subscription, or periodically retrieved via historical database queries to avoid missing events due to network fluctuations.

[0033] Maintenance and repair records can be obtained from the enterprise's equipment management system, inspection system, EAM system, CMMS system, or manual data entry terminals. Maintenance and repair records must include at least the inspection time, repair time, replaced parts, maintenance measures, maintenance completion time, maintenance results, and maintenance responsibility information. For scenarios not yet connected to a digital operation and maintenance system, data can be supplemented through manual entry templates, mobile inspection terminals, or electronic work order systems, and then synchronized to the health management platform via a data interface.

[0034] Environmental data is acquired from station environmental monitoring devices, meteorological sensors, or utility monitoring systems. Environmental data may include at least ambient temperature, humidity, station wind speed, and ambient pressure; when necessary, external environmental quantities such as cooling medium supply conditions and external power fluctuations may also be collected. Environmental data is used to eliminate the impact of environmental disturbances on the judgment of compressor unit operating status; therefore, it is collected or mapped synchronously with operating data.

[0035] Historical fault sample data is extracted from historical databases, accident record databases, downtime analysis report databases, maintenance archives, or existing fault diagnosis systems. Historical fault sample data includes operational data fragments from a period prior to the fault occurrence, corresponding control command sequences, alarm event sequences, maintenance and repair records, and the final fault conclusion. To facilitate subsequent sequence similarity comparison and causal propagation analysis, the historical fault sample data undergoes unified time alignment, field standardization, and fault tag archiving.

[0036] Regarding the acquisition cycle setting, it is set differently according to the data type, rather than using a uniform single cycle for all signals. For continuous process quantities such as pressure, temperature, flow rate, anti-surge valve opening, current or power, lubrication-related parameters, and cooling-related parameters, acquisition cycles of 1 to 10 seconds are used. Furthermore, under typical station monitoring conditions, fixed acquisition cycles of 1 second, 2 seconds, 5 seconds, or 10 seconds can be used, with 5 seconds being a commonly used implementation that balances real-time performance and storage costs. For speed signals, acquisition cycles of 1 to 5 seconds are used; when speed fluctuations are rapid or when monitoring of start-stop transient processes is required, the cycle can be shortened to 1 second. For vibration data, if the original waveform is acquired, it is locally acquired at millisecond or higher frequencies using a high-frequency sampling device, and features such as vibration amplitude, frequency band energy, peak value, and root mean square value are extracted at the edge side before being uploaded at a cycle of 1 to 10 seconds. If only vibration characteristic quantities are collected, they can be uploaded directly at a cycle of 1 to 5 seconds. For control command data and alarm event data, a combination of event-triggered acquisition and timed polling acquisition is used. Event-triggered acquisition can record immediately when a command is issued or an alarm occurs, while the timed polling acquisition cycle is 1 to 5 seconds.

[0037] For maintenance and repair records and environmental data, maintenance and repair records can be recorded when maintenance actions occur, end, or are confirmed. Environmental data is collected at a cycle of 10 to 60 seconds; in scenarios where environmental changes are slow, a 1-minute cycle can also be used. For historical fault sample data, real-time collection is usually not required; instead, it is imported in batches by fault event dimension through offline extraction.

[0038] The system employs the following data acquisition rhythm: pressure, temperature, flow rate, valve position, current or power, lubrication parameters, and cooling parameters are acquired at 5-second intervals; rotational speed is acquired at 1-second intervals; raw vibration data is sampled at high frequency on the edge side, and vibration characteristic values ​​are output every 1 second; control commands and alarm events are recorded using event triggering with 1-second polling verification; environmental data is acquired at 30-second intervals; maintenance records are written immediately after maintenance actions are completed; historical fault sample data is imported in batches according to fault events. This layered acquisition method ensures that critical changes in the compressor unit's status are captured promptly while avoiding excessive communication and storage pressure caused by uploading all data at high frequencies.

[0039] To ensure the accuracy of subsequent health assessments and fault predictions, the acquisition timestamp, source identifier, device identifier, and acquisition status identifier of each data point are further recorded after data acquisition. When there are missing, delayed, or outlier values ​​within a certain acquisition period, missing, continuity, and validity markers are generated synchronously for subsequent dynamic valid signal subset filtering and missing pattern recognition.

[0040] Various preset thresholds, preset durations, and preset boundaries are jointly determined based on the historical valid operating data of the equipment group and the historical stable operating data of the target compressor unit. Specifically, for field integrity thresholds, time continuity thresholds, short missing thresholds, stable operating condition segment judgment thresholds, reference boundaries, and risk change boundaries, it is preferable to first extract historical stable samples under the same operating condition anchor point and the same or similar load range, and then statistically analyze the center value, quantile interval, upper limit of fluctuation, and lower limit of fluctuation of the corresponding indicators. Then, individual corrections are performed in conjunction with the single-unit correction vector of the target compressor unit to obtain preset thresholds or boundaries applicable to the current target compressor unit. For preset recovery durations, preset expected values, and suggested execution windows, it is preferable to obtain them based on the abnormal duration, maintenance recovery time, and risk fallback time statistics in historical fault samples.

[0041] To avoid a decline in adaptability due to a fixed threshold over a long period, it is preferable to update the data according to a preset update cycle or after the cumulative number of newly added valid samples reaches a preset number, but the threshold and boundary should remain unchanged within a single monitoring cycle.

[0042] Based on the time synchronization tolerance window, start / stop state switching time, load interval switching time, anti-surge action time, and maintenance action time, the system performs time alignment and operating condition anchor point mapping on operating data, control command data, alarm event data, and maintenance records to generate a unified operating condition anchor point reference axis and a unified event record sequence arranged according to it.

[0043] Specifically, a standard time series formed by a preset collection period is used as the basic index. The preset collection period can be set to any one of 1 second, 2 seconds, 5 seconds, 10 seconds or other fixed periods according to the monitoring accuracy requirements. In order to eliminate the recording deviation between different data sources, a time synchronization tolerance window is set. When the time difference between the original timestamp of a data record and a certain basic index point does not exceed the time synchronization tolerance window, the data record is merged into the corresponding basic index point. When it exceeds the time synchronization tolerance window, delayed matching, interpolation compensation or missing mark processing is performed according to the data type.

[0044] For continuously sampled data, resampling mapping is performed according to the basic index; when the original sampling frequency is higher than the basic index frequency, mean, weighted mean, median, last value preservation or feature extraction are performed on multiple sampled values ​​within the same index interval to obtain the resampled value corresponding to that index point; When the original sampling frequency is lower than the basic index frequency, linear interpolation, step preservation, or spline interpolation is used to generate supplementary values ​​between adjacent valid sampled values. When no valid value is available for interpolation or the interval between adjacent valid values ​​exceeds the preset allowable span, the corresponding field is marked as missing. For vibration data, if the original waveform is collected, the root mean square value, peak value, frequency band energy, or kurtosis, etc., are extracted from the edge side first, and then mapped according to the basic index.

[0045] For control command data, it is mapped to the base index according to the command's effective time. For instantaneous control commands, their timestamps are aligned to the nearest base index point, and the command type, command value, and command source are written into the corresponding unified event record. For control states with continuous effects, the effective range is determined based on the state's start and end times, and all base index points covered by this range are marked as valid for the corresponding control state. For alarm event data, it is mapped to the base index according to the trigger time, release time, or duration range. For persistent alarms, all basic index points covered between the trigger and deactivation times are marked as alarm valid, and an alarm recovery flag is written to the first basic index point after the alarm is deactivated. For maintenance records, the basic index is mapped according to the start time of the action, the end time of the action, and the recovery observation interval. Among them, the basic index points covered from the start time of maintenance to the end time of maintenance are marked as the maintenance action interval, and the basic index points covered by the preset observation period after maintenance are marked as the maintenance recovery phase. If a certain inspection and maintenance record only contains the completion time, then the completion time is used as the anchor point, and the action range is added or marked as an instantaneous maintenance event according to the preset forward backtracking window.

[0046] After completing time alignment, further execution of operating condition anchor point mapping is performed. Specifically, at least start-stop phase anchor points, load interval anchor points, anti-surge action status anchor points, and maintenance and recovery phase anchor points are set. Start-stop phase anchor points can be based on the time of start command issuance, the time when the speed rises to a preset threshold, the time of stable operation start, the time of stop command issuance, and the time when the speed drops to zero or close to zero, dividing the operation process into start preparation phase, start transition phase, stable operation phase, stop transition phase, and stop completion phase; load interval anchor points can be based on flow rate, power, load rate, or pressure ratio to divide the stable operation phase into low load interval, medium load interval, and high load interval, or into multiple discrete load levels; The anti-surge action status anchor point can be marked as normal anti-surge state, anti-surge adjustment state, active reflux state or anti-surge recovery state according to the change of anti-surge valve opening, anti-surge control command and related protection logic; The anchor points for the maintenance and recovery phase can be marked as short-term recovery phase, medium-term recovery phase, or recovery completion phase based on the maintenance completion time and recovery observation time window.

[0047] Anchor points for various operating conditions are written to a unified event record using hierarchical or combined marking methods, so that each basic index point has both time location and running stage attributes.

[0048] Based on this, a unified operating condition anchor point reference axis is generated. The unified operating condition anchor point reference axis is formed by combining a basic time index and operating condition anchor points. Each index unit includes at least a basic time point and one or more of the following: start / stop phase marker, load interval marker, anti-surge action status marker, and maintenance / recovery phase marker. Subsequently, for each basic index point, the operational data, control command data, alarm event data, and maintenance records that have completed time alignment and operating condition anchor point mapping are summarized to form a unified event record. The unified event log includes at least the device identifier, basic index time, operating condition anchor point identifier, operating parameter field, control status field, alarm status field, maintenance status field, field validity flag, missing flag, source identifier, and quality flag.

[0049] For fields with multiple source values ​​or conflicting values ​​at the same base index point, conflict resolution is performed based on data source priority, time proximity, and field quality markers; all unified event records are arranged in the order of the base time index to obtain the unified event record sequence.

[0050] After obtaining a unified event record sequence, a mechanism reference spectrum is constructed only under the condition of the same operating condition anchor point and in the stable operating condition segment.

[0051] The stable operating condition can be identified based on one or more of the following conditions: the rate of change of speed is lower than the preset threshold, the flow fluctuation amplitude is lower than the preset threshold, the rate of change of power is lower than the preset threshold, the anti-surge valve opening is maintained in the stable range, the control command has not changed for a new time within the preset duration, and the alarm status has not entered a high-level abnormal state. When several consecutive basic index points simultaneously meet the above conditions, the consecutive segment is marked as a stable operating segment.

[0052] The mechanism reference spectrum is constructed based on stable segment samples from the historical valid operating data of the equipment group that are at the same operating condition anchor point, the same or similar load range as the current target compressor unit, and whose data quality meets the requirements. It includes at least the correlation constraint relationship between pressure ratio and flow rate, the correlation constraint relationship between speed and power, the correlation constraint relationship between anti-surge valve opening and flow fluctuation, and the correlation constraint relationship between exhaust temperature and load. Association constraints can be expressed using table ranges, fitted curves, piecewise functions, regression models, envelope boundaries, or multidimensional lookup tables.

[0053] To eliminate the impact of individual equipment differences, a single-unit correction vector can be generated based on the long-term offset of the target compressor unit relative to the equipment group reference relationship during its historical stable operation phase, and individual corrections can be made when calling the mechanism reference spectrum.

[0054] After obtaining the mechanism reference spectrum, a causal feasible region is further established. Specifically, a candidate causal edge is allowed only when a preceding node and a subsequent node simultaneously satisfy the following conditions: the change of the variable corresponding to the preceding node can change before the change of the subsequent node in the compressor unit operation mechanism, satisfying the physical propagation order of the compressor unit; The variables corresponding to the preceding and following nodes have verifiable correlation links in the mechanistic reference spectrum, satisfying the variable coupling closure relationship; The preceding and following nodes are located under the same working condition anchor point, or under adjacent working condition anchor points that allow propagation, satisfying the working condition anchor point consistency requirement.

[0055] By employing the above method, candidate edges in the initial precursor transition graph are subject to pre-constraints of operating mechanism, variable correlation, and working condition context before entering subsequent causal identification, thereby reducing spurious related edges and improving the reliability of subsequent causal analysis and critical path identification.

[0056] First, stable operating conditions are identified in the unified event log sequence, and a mechanism reference spectrum is constructed based on the historical valid operating data of the equipment group under the same operating condition anchor point.

[0057] Specifically, under the same operating conditions and anchor points, a sliding window method is used to analyze the unified event record sequence segment by segment. Within each analysis window, the speed change rate, flow fluctuation amplitude, power change rate, pressure ratio fluctuation amplitude, anti-surge valve opening change rate, control command jump number, and alarm status change are calculated respectively. When the rate of change of rotational speed, the amplitude of flow fluctuation, the rate of change of power, and the rate of change of anti-surge valve opening do not exceed the corresponding thresholds, and there are no new start-stop control commands, load switching commands, anti-surge switching commands, or high-level alarms within the window, the window is determined as a stable candidate window. When the operating condition anchor points of multiple consecutive stable candidate windows are consistent or belong to adjacent anchor point intervals that are allowed to be merged, they are merged into a stable operating condition segment.

[0058] Subsequently, compressor unit samples of the same type as the target compressor unit or meeting the preset similar conditions are selected from the historical database of the equipment group. Then, historical stable operating conditions that are consistent with the current operating condition anchor point, have the same or adjacent load range, the same or similar medium conditions, and meet the data quality requirements are extracted as modeling samples. Samples with obvious faults, maintenance actions in progress, frequent switching of control modes, large-scale missing key fields, or continuously high alarm levels are removed. Then, according to the operating condition anchor point and load level, the normal correlation between pressure ratio and flow rate, speed and power, anti-surge valve opening and flow fluctuation, and exhaust temperature and load are statistically analyzed. The reference center value, allowable fluctuation range, upper boundary and lower boundary under the corresponding operating condition anchor point are output in any one or more forms, such as table interval, quantile envelope, fitted curve, piecewise function, regression model, multidimensional lookup table or reference surface, thereby forming the mechanism reference spectrum.

[0059] To eliminate the impact of individual equipment differences, samples that are consistent with the current operating condition anchor point and have been identified as stable operating conditions are further selected from the historical effective operating data of the target compressor unit. These samples are compared with the corresponding mechanism reference spectrum of the equipment group in each key variable dimension. One or more of the following are calculated: pressure ratio offset, flow rate offset, power offset, anti-surge valve opening offset, and exhaust temperature offset. The mean offset, median offset, quantile offset, or weighted moving average offset are used as the single-unit offset components of each variable. The offset components of each variable are combined in a preset order to form a single-unit correction vector.

[0060] Based on the mechanism reference spectrum, the causal feasible region is further limited. Candidate causal edges are only allowed to be established when a certain preceding node and the following node simultaneously satisfy the requirements of the physical propagation order of the compressor unit, the closed relationship of variable coupling, the consistency of the operating condition anchor point, and the preset time lag range. This ensures that subsequent causal identification is carried out only within the set of candidate edges constrained by the mechanism.

[0061] Based on this, a subset of dynamically valid signals is selected according to the missing status, continuity status, and correlation closure status of each signal within the current monitoring period, and the missing pattern corresponding to the current monitoring period is identified. Specifically, for each operating signal within the current monitoring period, field integrity index, time continuity index, and correlation closure index are calculated. The field integrity index is determined by the ratio of the number of valid sampling points to the number of required sampling points. The time continuity index is jointly determined by the maximum interval, average interval, consecutive missing length, and number of breakpoints between adjacent valid sampling points. The correlation closure index is determined by whether a closed relationship can still be formed between the signal and the core mechanism variable, which can be used for residual calculation and propagation analysis. When a certain operating signal has a field integrity index that is not lower than the first threshold, a time continuity index that is not lower than the second threshold, and satisfies the correlation closure condition with at least two core mechanism variables, it is included in the dynamic effective signal subset, where the core mechanism variables include at least pressure ratio, flow rate, speed, power, and anti-surge valve opening.

[0062] For operating signals that, despite having local defects, can still be reliably corrected using adjacent effective values, mechanistic reference spectra, and single-machine correction vectors, they can be included in the dynamic effective signal subset after adding a quality label; For operating signals that are continuously missing for too long, have too many breakpoints, or have lost their closed relationship with core mechanism variables, they are removed from the dynamic valid signal subset, and only their missing markers are retained for subsequent missing pattern identification. Missing patterns are determined by jointly judging the distribution of missing locations, the duration of missing data, the importance level of missing variables, and the number of missing sources: when missing points are discretely distributed, the length of a single continuous missing data does not exceed the preset short missing threshold, and they do not occur concentratedly on core mechanism variables, they are judged as random missing; when an operating signal or several operating signals are continuously missing at multiple consecutive basic index points, the length of continuous missing data exceeds the preset short missing threshold, and forms obvious missing segments, they are judged as continuous segment missing; when the missing signal belongs to one or more of the pressure ratio, flow rate, speed, power, and anti-surge valve opening, or although the number of missing signals is small, it has caused the corresponding residual term to be unable to be calculated in the normal way, it is judged as a critical variable missing. When multiple signals from different data sources are missing simultaneously in the same or highly overlapping time intervals, and involve two or more types of data sources such as process quantities, control quantities, or event quantities, it is determined to be a synchronous multi-source missing signal. The corresponding missing mode identifier and the identifier of the dynamic valid signal subset are written into the current monitoring cycle status record so that normal prediction calculation and degraded causal chain reconstruction are carried out in the same residual expression space.

[0063] After determining the dynamic effective signal subset, the equivalent residual packet is calculated based on the dynamic effective signal subset, the mechanism reference spectrum and the single-machine correction vector. Health factors are extracted based on short time windows and long time windows. Then, the residual-health coupling state matrix is ​​constructed based on the joint change of the equivalent residual packet and the health factors.

[0064] Specifically, for each operating signal included in the dynamic effective signal subset, the reference value or reference range corresponding to the signal is first read from the mechanism reference spectrum based on the current operating condition anchor point, load range and corresponding variable combination. Then, the reference value or reference range is individually corrected by combining the single unit correction vector to obtain the corrected reference value or corrected reference range of the target compressor unit under the current operating condition. Subsequently, the actual measured values ​​within the current monitoring period are compared with the corrected reference values ​​to obtain the corresponding residual terms.

[0065] For variables with a one-to-one mapping relationship, the difference between the measured value and the corrected reference value, the absolute deviation, the relative deviation, or the normalized deviation are used as the residual terms. For a state quantity characterized by multiple variables, a combined residual term is generated based on the degree of deviation of the corresponding variable combination from the mechanism reference spectrum.

[0066] The equivalent residual package includes at least the pressure ratio residual, flow rate residual, power residual, exhaust temperature residual, vibration deviation, and steady-state recovery residual after control execution. The steady-state recovery residual after control execution can be determined by comparing the actual recovery value within the preset stable time after the control action is completed with the corrected reference recovery value.

[0067] Health factors include at least alarm co-occurrence density, control response lag, start-stop disturbance degree, maintenance recovery rate, cross-source consistency score, and missing continuity index. Among them, alarm co-occurrence density can be determined by weighting the number, duration, and level of alarm events within the window; control response lag can be determined by the time difference between the issuance of control commands and the achievement of preset response conditions by key process quantities; start-stop disturbance degree can be determined by comprehensively considering the deviation amplitude and fluctuation duration of process quantities within the start-stop phase; maintenance recovery rate can be determined by the proportion or speed at which key variables regress to the corrected reference value within the maintenance recovery phase; cross-source consistency score can be determined by the consistency of indications from multiple data sources for the same operating state; and missing continuity index can be determined by the continuous length, distribution density, and degree of influence on key variables of missing points within the window.

[0068] Each health factor is preferably converted into a comparable standardized quantity before participating in the coupling analysis. Specifically, the alarm co-occurrence density is preferably obtained by weighted summation and normalization based on the number of alarms, duration, and alarm level within the window; the control response hysteresis is preferably obtained by normalization based on the time difference between the time of control command issuance and the time when the key process quantity reaches the preset response condition. The optimal start-stop disturbance level is calculated based on the combined calculation of the maximum deviation of key process quantities, the duration of deviation, and the number of fluctuations during the start-stop phase. The optimal maintenance recovery rate is calculated by combining the proportion and the rate of regression of the target variable to the corrected reference value after maintenance. Cross-source consistency score optimization is obtained based on the consistency of multiple data sources' judgments on the same running state; The missing continuity index is preferably calculated by combining the length of consecutive missing values, the density of missing values, and the importance level of the variables involved.

[0069] Ideally, all health factors should be mapped to the same numerical range, and the direction should be made consistent so that the larger the factor value, the higher the degree of abnormality or the higher the degree of health deterioration.

[0070] Furthermore, under the same operating condition anchor point, the residual terms in the equivalent residual package and the factor terms in the health factor are cross-combined according to short-time windows and long-time windows to form a residual-health coupled state matrix. The short-time window is used to characterize rapid abnormal fluctuations and control response lag, while the long-time window is used to characterize the continuous degradation trend and the decay trend of maintenance and recovery effect. For each state unit, the current value, rate of change, and cumulative offset of the corresponding residual term, as well as the current value, rate of change, and duration of the corresponding health factor are recorded to characterize the offset level, evolution direction, and persistence of the state unit.

[0071] The residual-health coupled state matrix is ​​preferably generated using "residual item - health factor - operating condition anchor point - time window" as the state unit index. Specifically, for each time window under each operating condition anchor point, each residual item in the equivalent residual package is paired with each factor item in the health factor to form multiple state units. Each state unit records at least the current value of the residual item, the residual change rate, the cumulative offset, the current value of the corresponding health factor, the health factor change rate, the duration, and the current operating condition anchor point identifier. When the residual item and the health factor in a certain state unit simultaneously exceed their respective reference boundaries within the same time window, or when the residual item crosses the boundary first and the health factor deteriorates within a preset lag range, it is determined to be a synchronous deterioration state. When only the residual item crosses the boundary and the health factor does not deteriorate accordingly, it is preferably marked as a state to be confirmed and is not directly used for the generation of degraded segments.

[0072] In this embodiment, each residual term in the equivalent residual package is preferably generated according to a uniform rule. Specifically, the pressure ratio residual is preferably the difference between the current measured pressure ratio and the corrected reference pressure ratio at the corresponding operating condition anchor point, or the normalized deviation of the difference relative to the corrected reference pressure ratio; the flow rate residual, power residual, and exhaust temperature residual are calculated in the same way. The vibration deviation is preferably the absolute or normalized deviation of the current vibration characteristic value relative to the corresponding corrected reference vibration characteristic value; the steady-state recovery residual after control execution is preferably the difference between the average actual recovery value and the average corrected reference recovery value within the preset stable observation interval after the control action is completed. For state variables jointly characterized by multiple variables, it is preferable to first calculate the individual variable residuals separately, and then combine them according to preset weights to obtain combined residual terms; the residual terms are arranged in a fixed order to form an equivalent residual package to ensure that normal prediction calculation and downgraded causal chain reconstruction call the same residual expression structure.

[0073] After obtaining the residual-health coupling state matrix, when the residual-health coupling state matrix undergoes a cross-boundary change relative to the reference boundary of the corresponding stable segment and maintains monotonic evolution within a long time window, a degenerate segment is generated, and the segment confidence score, coupling state identifier, and precursor feature vector are determined.

[0074] The reference boundary is preferably generated statistically based on historical stable segment samples under the same operating condition anchor point and the same or similar load range, and includes at least the center value, allowable fluctuation range, upper boundary and lower boundary of the corresponding state unit; If necessary, the center value and boundary can be corrected based on the single-unit correction vector of the target compressor unit.

[0075] After the degraded fragment is generated, the fragment credibility score is preferably generated based on the field integrity, temporal continuity, cross-source consistency, working condition anchor stability, and causal feasible domain conformity within the fragment segment. Then, the scores are weighted according to preset weights to obtain the total score. Preferably, the total score is divided into three levels: high credibility, medium credibility, and low credibility, for use in subsequent precursor transition graph construction, path pruning, and suggested work order generation.

[0076] Specifically, the reference boundary is determined based on historical stable segment samples under the same operating condition anchor point. It can be constructed based on the center value, allowable fluctuation range, upper boundary, and lower boundary of each state unit in the stable state. When the residual term, health factor, or a combination of both of a state unit deviates beyond the corresponding reference boundary, it is determined that the state unit has undergone cross-boundary change.

[0077] To avoid misjudgments caused by transient noise or short-term disturbances, the corresponding continuous segment is marked as a degenerate segment only when the cross-boundary change meets the following conditions within a long time window: consistent evolution direction, duration exceeding a preset duration, and no conflict with the causal feasible region. Consistent evolution direction means that the corresponding residual term and health factor maintain the same deterioration trend or maintain a trend consistent with the preset degeneracy direction at multiple consecutive basic index points. Duration exceeding a preset duration means that the duration of the cross-boundary state is not less than a preset threshold. No conflict with the causal feasible region means that the variable propagation relationship corresponding to the cross-boundary change does not violate the requirements of the physical propagation order of the compressor unit, the variable coupling closure relationship, and the consistency of the operating condition anchor point.

[0078] The starting position of the degenerate segment is the basic index point that first satisfies the cross-boundary change and the condition of persistence, and the ending position is the basic index point where the state unit falls back into the reference boundary and remains stable for more than the preset recovery time.

[0079] The credibility score of a fragment is determined by comprehensively considering the completeness of fields, temporal continuity, cross-source consistency, stability of anchor points, and compliance with causal feasible regions within the degraded fragment segment. The coupling state identifier is determined based on the combination type of the main residual terms and main health factors within the degraded segment; The precursor feature vector is formed by combining the main residual statistics, main health factor statistics, cross-boundary amplitude, duration, change slope, recovery speed, operating condition anchor point identifier, and segment confidence score within the degraded segment. It is used to characterize the abnormal nature and evolutionary characteristics of the degraded segment.

[0080] After obtaining multiple degraded segments, an initial precursor transition graph is constructed based on the temporal adjacency relationship, the continuity relationship of the working condition anchor point, the consistency relationship of the residual direction, the maintenance of reversibility, the multi-source support relationship, and the causal feasible region among the degraded segments. Then, under the same working condition anchor point, causal direction identification, confusion factor detection, and path pruning are performed on the candidate edges to generate a causal precursor transition graph. Finally, the causal strength coefficient and causal edge credibility are determined for the retained causal edges.

[0081] Specifically, for any two degenerate segments, if the start time of the subsequent degenerate segment is later than the end time of the preceding degenerate segment or is within the allowable time lag range, and both are under the same working condition anchor point or under adjacent working condition anchor points that are allowed to propagate, then the temporal adjacency relationship and the working condition anchor point continuity relationship are considered to be satisfied. If the main residual terms of the two segments change in the same direction, or if the abnormal direction of the preceding degenerate segment can be guided by the abnormal direction of the subsequent degenerate segment in terms of mechanism, then the relationship of consistent residual direction is considered to be satisfied. If the preceding degraded segment shows a significant decline after inspection and maintenance and the subsequent degraded segment weakens or disappears synchronously, then the maintenance reversibility relationship is considered to be satisfied. If multiple data sources provide consistent support for the existence and evolution order of two consecutive degenerate segments, then the multi-source support relationship is considered to be satisfied. Furthermore, candidate edges are only allowed to be established between two degenerate segments if the relationship between their corresponding variables satisfies the causal feasible region constraint.

[0082] For the established candidate edges, for any two nodes A and B in the initial precursor transition graph that have a candidate edge relationship, a time series causal test model based on the historical values ​​of node A and node B is constructed under the same working condition anchor point to determine whether the historical value of node A has a significant contribution to the prediction result of node B. The transfer entropy from node A to node B and the transfer entropy from node B to node A are calculated respectively. When the transfer entropy from node A to node B is greater than the transfer entropy from node B to node A and the difference between the two exceeds a preset threshold, and the time series causal test model supports that node A has a significant impact on node B, the direction from node A to node B is determined to be the dominant causal direction and the corresponding candidate edge is retained.

[0083] The confusion factor detection selects at least one or more of the following as potential confusion factor candidates: anchor point type, ambient temperature, load level, and cumulative equipment runtime. Under the given potential confusion factor candidates, conditional independence tests are performed on the preceding and following nodes corresponding to the candidate edges. For candidate edges retained after conditional independence testing, initial causal edge credibility is generated based on causal direction identification results, propagation entropy direction advantage, working condition anchor point consistency, multi-source support degree, and mechanistic reference spectrum constraint compliance. Each retained edge is then weighted using a causal strength coefficient. Furthermore, the dominant causal transfer sequence is jointly extracted based on path cumulative weight, causal strength coefficient, causal edge credibility, and sequence similarity with historical fault sample data at the same working condition anchor point. The optimal blocking point scheme is determined comprehensively based on the risk reduction, implementation cost, implementation difficulty, and estimated recovery benefit of each node or edge in the key causal propagation path.

[0084] The initial causal edge credibility is preferably generated using a weighted multi-score method, where the causal direction identification result, propagation entropy direction advantage, working condition anchor point consistency, multi-source support degree, and mechanism reference spectrum constraint compliance degree each correspond to a single-item score. The credibility of the causal edge is obtained by summing the normalized scores according to preset weights. The optimal blocking point scheme preferably calculates the risk reduction, implementation cost, implementation difficulty, and recovery benefit estimate after blocking for each node or edge on the key causal propagation path, and converts these into comparable scores. Among them, the risk reduction amount and the estimated recovery benefit are positive indicators, while the implementation cost and implementation difficulty are negative indicators. After directional consistency processing, they are weighted according to preset weights to obtain the candidate blocking point scores. The one with the highest score is selected as the optimal blocking point, and the one with the second highest score can be used as the alternative blocking point.

[0085] The preliminary selection of potential confounding factors is first conducted based on the type of operating condition anchor point, ambient temperature, load level, cumulative equipment runtime, and, if necessary, media conditions, cooling conditions, and maintenance status. Then, factors that are statistically correlated with both node A and node B and have process significance are selected as inputs for the conditional independence test.

[0086] If, given a certain potential confusion factor or a combination of potential confusion factors, the correlation between node A and node B drops below a preset independence threshold, or its causal direction test no longer meets the preset significance level, then the corresponding candidate edge is deleted; if, given a confusion factor, the directional significance and transit entropy advantage are still met, then the candidate edge is retained and enters the subsequent causal edge credibility calculation.

[0087] When the dynamic effective signal subset does not meet the normal prediction calculation conditions, a degraded causal chain reconstruction is performed based on the missing mode to generate a degraded causal precursor transition graph.

[0088] Specifically, when the missing pattern is random missing, local residual compensation is performed on the missing location based on adjacent valid sampling points and adjacent valid operating condition anchor points, while keeping the cause-effect precursor transition map topology unchanged. When the missing pattern is a continuous segment missing, a bridging causal edge is constructed based on the stable operating segment before and after the missing segment, the coupling state identifier before and after the missing segment, and the dominant causal transition sequence of the previous effective monitoring cycle. When the missing mode is the absence of a key variable, a substitute variable that satisfies the mechanistic closure relationship with the missing variable is selected. An equivalent bridging residual is generated by combining the mechanistic reference spectrum and the single-machine correction vector, and the key causal chain is reconstructed based on the equivalent bridging residual. When the missing mode is the absence of multiple sources simultaneously, a conservative update is performed based on the remaining dynamic effective signal subset, the dominant causal transfer sequence of the previous effective monitoring period, and the preset risk change boundary, and the recommended execution window is shortened.

[0089] Based on the causal precursor transition map or the downgraded causal precursor transition map, the dominant causal transition sequence, key causal propagation paths, and optimal blocking point solutions are extracted. A health index, fault risk score, and risk status label are generated. Based on the optimal blocking point solution, the estimated recovery benefit, and the suggested execution window, a suggested work order is generated. After the suggested work order is executed, intervention feedback data is collected. The intervention feedback data includes at least the risk reduction, alarm fading time, parameter regression degree, recovery time, and actual intervention cost. Based on the intervention feedback data, the causal edge credibility, causal strength coefficient, fault rule template, and optimal blocking point solution scoring parameters are self-corrected and updated.

[0090] Specifically, for each causal edge on the key causal propagation path, the risk reduction, alarm fading time, parameter regression degree, recovery time, and actual intervention cost after the corresponding intervention measures are implemented are statistically analyzed. When the actual risk reduction after the intervention measures corresponding to a certain causal edge are implemented is greater than the preset expected value and the parameter regression degree meets the preset conditions, the causal edge credibility of that causal edge and the score of the corresponding node or edge in the optimal blocking point scheme are increased. When the actual risk reduction after the intervention measures corresponding to a certain causal edge are implemented is lower than the preset expected value or the recovery time exceeds the preset time, the causal edge credibility of that causal edge is decreased, and the weight of the corresponding fault rule template and the score of the optimal blocking point scheme are lowered, thereby outputting health management results that meet the risk continuity constraint and physical consistency constraint.

[0091] A process equipment failure prediction and health management system, comprising: The data access module is used to acquire the operating data of the target compressor unit equipment group in the oil and gas business according to the preset acquisition cycle, and simultaneously acquire control command data, alarm event data, maintenance records, environmental data and historical fault sample data. The working condition anchor point mapping module is used to generate a unified working condition anchor point reference axis and output a unified event record sequence; The mechanism reference construction module is used to construct the equipment group-level mechanism reference spectrum, generate the single-unit correction vector of the target compressor unit, and define the causal feasible region under the conditions of the same operating condition anchor point and in the stable operating condition segment. The dynamic valid signal filtering and missing pattern identification module is used to filter a subset of dynamic valid signals and identify missing patterns based on field completeness, temporal continuity and correlation closure. The equivalent residual packet generation module is used to generate equivalent residual packets based on the dynamic effective signal subset, the mechanism reference spectrum, and the single-machine correction vector. The coupling state analysis module is used to generate a residual-health coupling state matrix based on the equivalent residual package and health factors and to identify degenerate segments. The initial precursor transition graph construction module is used to construct the initial precursor transition graph based on the temporal adjacency relationship between degenerate segments, the continuity relationship of operating condition anchor points, the consistency relationship of residual direction, the maintenance of reversibility relationship, the multi-source support relationship, and the causal feasible region. The causal pruning and critical path identification module is used to perform causal direction identification, confusion factor detection, conditional independence test, causal strength coefficient calculation, causal edge confidence determination, path pruning, and extraction of dominant causal transfer sequence and optimal blocking point scheme under the same working condition anchor point. The downgraded causal chain reconstruction module is used to perform downgraded causal chain reconstruction based on the missing mode and generate a downgraded causal precursor transition map when the dynamic effective signal subset does not meet the normal prediction calculation conditions. The risk fusion assessment module is used to generate health index, fault risk score, risk status label and suggested work order; The intervention feedback self-correction module is used to update the causal edge confidence, causal strength coefficient, fault rule template weight, and optimal blocking point scheme score based on intervention feedback data after the suggested work order is executed. The system adopts an online hybrid deployment approach, including at least a field edge processing unit and a central analysis and processing unit. The field edge processing unit is used to perform data access, dynamic effective signal filtering, missing pattern identification, downgraded causal chain reconstruction, and conservative update calculation. The central analysis and processing unit is used to perform device group-level mechanism reference spectrum update, causal pruning and critical path identification, intervention feedback self-correction, and risk fusion assessment.

[0092] In conclusion, the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for predicting and managing the health of process equipment failures, characterized in that, include: According to the preset acquisition cycle, the operating data of the target compressor unit is acquired, and control command data, alarm event data, maintenance records, environmental data and historical fault sample data are acquired simultaneously. The operating data includes at least inlet pressure, outlet pressure, inlet temperature, exhaust temperature, flow rate, vibration, speed, current or power, anti-surge valve opening, lubrication-related parameters and cooling-related parameters. Based on the time synchronization tolerance window, start-stop state switching time, load interval switching time, anti-surge action time, and maintenance action time, the operation data, control command data, alarm event data, and maintenance records are time aligned and operating condition anchor points are mapped to generate a unified operating condition anchor point reference axis and a unified event record sequence arranged according to it. In the unified event record sequence, stable operating segments are identified, and a mechanism reference spectrum is constructed based on the historical effective operating data of the equipment group under the same operating condition anchor point. At the same time, the single-unit correction vector is determined based on the historical offset of the target compressor unit relative to the equipment group, and the causal feasible region between the preceding and following nodes that satisfies the physical propagation order of the compressor unit is defined by the mechanism reference spectrum. Based on the missing status, continuity status, and correlation closure status of each signal within the current monitoring period, a subset of dynamically valid signals is selected from the operational data, and the missing pattern corresponding to the current monitoring period is identified. The equivalent residual packet is calculated based on the dynamic effective signal subset, mechanism reference spectrum and single-machine correction vector. Health factors are extracted based on short time windows and long time windows. The residual-health coupling state matrix is ​​constructed based on the joint change of the equivalent residual packet and health factors. When the residual-health coupling state matrix undergoes a cross-boundary change relative to the reference boundary of the corresponding stable segment and remains monotonically evolving within a long time window, a degenerate segment is generated, and the segment confidence score, coupling state identifier, and precursor feature vector are determined. An initial precursor transition graph is constructed based on the temporal adjacency relationship between degraded segments, the continuity relationship of working condition anchor points, the consistency relationship of residual direction, the maintenance of reversibility, the multi-source support relationship, and the causal feasible region. For candidate edges in the initial precursor transition graph, causal direction identification, confusion factor detection and path pruning are performed under the same working condition anchor point to generate a causal precursor transition graph, and the causal strength coefficient and causal edge confidence are determined for the retained causal edges. When the dynamic effective signal subset does not meet the normal prediction calculation conditions, a degraded causal chain reconstruction is performed based on the missing mode to generate a degraded causal precursor transition map. Based on the causal precursor transition map or the downgraded causal precursor transition map, the dominant causal transition sequence, key causal propagation paths and optimal blocking point solutions are extracted, and a health index, fault risk score and risk status label are generated. Based on the optimal blocking point solution, the recovery benefit estimate and the suggested execution window, a suggested work order is generated. After the recommended work order is executed, intervention feedback data is collected, and self-correction updates are performed on the causal edge credibility, causal strength coefficient, fault rule template and optimal blocking point scheme scoring parameters based on the intervention feedback data, so as to output health management results that meet the risk continuity constraint and physical consistency constraint.

2. The method for predicting and managing the health of process equipment according to claim 1, characterized in that: The generation of the unified operating condition anchor point reference axis includes: using the standard time series formed by the preset acquisition cycle as the basic index, resampling the continuous sampling data to the basic index, mapping the alarm event data and maintenance records to the corresponding index segments according to the trigger time, release time or duration interval, and setting the compressor unit's start-up and shutdown phase, load interval, anti-surge action status and maintenance recovery phase as different types of operating condition anchor points respectively. The mechanism reference spectrum is constructed under the same operating condition anchor point and in the stable operating range, and includes at least the correlation constraint relationship between pressure ratio and flow rate, the correlation constraint relationship between speed and power, the correlation constraint relationship between anti-surge valve opening and flow fluctuation, and the correlation constraint relationship between exhaust temperature and load. The causal feasible region is limited by the mechanism reference spectrum to establish candidate causal edges only when the physical propagation sequence of the compressor unit, the variable coupling closure relationship, and the consistency of the operating condition anchor point are satisfied.

3. The method for predicting and managing the health of process equipment according to claim 1, characterized in that: The dynamic effective signal subset is determined by calculating the field integrity index, time continuity index, and correlation closure index of each operating signal in the current monitoring period. The operating signals with a field integrity index not lower than the first threshold, a time continuity index not lower than the second threshold, and a correlation closure condition with at least two core mechanism variables are included in the dynamic effective signal subset. Among them, the core mechanism variables include at least pressure ratio, flow rate, speed, power, and anti-surge valve opening; The missing pattern is obtained by jointly judging the distribution of missing locations, the duration of missing data, the importance level of missing variables, and the number of missing sources, so that normal prediction calculation and downgraded causal chain reconstruction can be carried out in the same residual expression space.

4. A method for predicting and managing the health of process equipment according to claim 1 or 3, characterized in that: The equivalent residual package includes at least the pressure ratio residual, flow rate residual, power residual, exhaust temperature residual, vibration deviation, and steady-state recovery residual after control execution; Health factors include at least alarm co-occurrence density, control response hysteresis, start-stop disturbance degree, maintenance recovery rate, cross-source consistency score, and missing continuity indicators; The residual-health coupled state matrix is ​​obtained by cross-combining the residual terms in the equivalent residual package with the factor terms in the health factor according to the operating condition anchor point and time window. The short time window is used to characterize rapid abnormal fluctuations and control response lag, while the long time window is used to characterize the continuous degradation trend and the decay trend of maintenance and recovery effect. The conditions for generating degenerate segments include: the residual-health coupling state matrix undergoes a cross-boundary change relative to the reference boundary of the corresponding stable segment, and within a long time window, it satisfies the following: consistent evolution direction, duration exceeding the preset duration, and no conflict with the causal feasible region.

5. The method and system for process equipment fault prediction and health management according to claim 3, characterized in that: Causal direction identification includes: for any two nodes A and B with candidate edge relationships in the initial precursor transition graph, a time series causal test model based on the historical values ​​of node A and node B is constructed under the same working condition anchor point to determine whether the historical value of node A has a significant contribution to the prediction result of node B, and the transfer entropy from node A to node B and the transfer entropy from node B to node A are calculated respectively. When the propagation entropy from node A to node B is greater than the propagation entropy from node B to node A and the difference between the two exceeds a preset threshold, and the time series causality test model supports that node A has a significant impact on node B, the direction from node A to node B is determined to be the dominant causal direction and the corresponding candidate edge is retained. The confusion factor detection selects at least one or more of the following as potential confusion factor candidates: anchor point type, ambient temperature, load level, and cumulative equipment runtime. Given the potential confusion factor candidates, conditional independence tests are performed on the preceding and following nodes corresponding to the candidate edges.

6. The method and system for predicting and managing process equipment failures according to claim 5, characterized in that: For candidate edges retained after the conditional independence test, the initial causal edge credibility is generated based on the causal direction identification result, the direction advantage of the transfer entropy, the consistency of the working condition anchor point, the degree of multi-source support and the constraint compliance of the mechanism reference spectrum, and the causal strength coefficient is combined to perform weighting on each retained edge. The dominant causal transition sequence is jointly extracted based on the cumulative path weight, causal strength coefficient, causal edge confidence, and sequence similarity with historical fault sample data under the same working condition anchor point. The optimal blocking point scheme is determined by comprehensively considering the risk reduction, implementation cost, implementation difficulty, and estimated recovery benefits of each node or side in the key causal propagation path.

7. The method and system for process equipment fault prediction and health management according to claim 3, characterized in that: Degraded causal chain reconstruction includes: when the missing pattern is random missing, while keeping the topology of the cause-effect precursor transition graph unchanged, performing local residual compensation on the missing location based on adjacent valid sampling points and adjacent valid operating condition anchor points; When the missing pattern is a continuous segment missing, a bridging causal edge is constructed based on the stable operating segment before and after the missing segment, the coupling state identifier before and after the missing segment, and the dominant causal transition sequence of the previous effective monitoring cycle. When the missing mode is the missing key variable, a substitute variable that satisfies the mechanistic closure relationship with the missing variable is selected, and an equivalent bridging residual is generated by combining the mechanistic reference spectrum and the single-machine correction vector. The key causal chain is then reconstructed based on the equivalent bridging residual. When the missing mode is synchronous multi-source missing, a conservative update is performed based on the remaining dynamic valid signal subset, the dominant causal transfer sequence of the previous valid monitoring period, and the preset risk change boundary, and the recommended execution window is shortened.

8. The method and system for predicting and managing process equipment failures according to claim 6, characterized in that: The self-correcting update of the credibility of causal edges includes: for each causal edge on the key causal propagation path, respectively, the risk reduction, alarm expiration time, parameter regression degree, recovery time and actual intervention cost after the corresponding intervention measures are implemented; When the actual risk reduction after the intervention measures corresponding to a certain causal edge are implemented are greater than the preset expected value and the parameter regression degree meets the preset conditions, the causal edge credibility of that causal edge and the score of the corresponding node or edge in the optimal blocking point scheme are increased. When the actual risk reduction after the intervention measures corresponding to a certain causal edge are implemented are lower than the preset expected value or the recovery time exceeds the preset time, the causal edge credibility of that causal edge is reduced, and the weight of the corresponding fault rule template and the score of the optimal blocking point scheme are lowered.

9. A process equipment fault prediction and health management system, characterized in that, include: The data access module is used to acquire the operating data of the target compressor unit equipment group in the oil and gas business according to the preset acquisition cycle, and simultaneously acquire control command data, alarm event data, maintenance records, environmental data and historical fault sample data. The working condition anchor point mapping module is used to generate a unified working condition anchor point reference axis and output a unified event record sequence. The mechanism reference construction module is used to construct the equipment group-level mechanism reference spectrum, generate the single-unit correction vector of the target compressor unit, and define the causal feasible region under the conditions of the same operating condition anchor point and in the stable operating condition segment. The dynamic valid signal filtering and missing pattern identification module is used to filter a subset of dynamic valid signals and identify missing patterns based on field completeness, temporal continuity and correlation closure. The equivalent residual packet generation module is used to generate equivalent residual packets based on the dynamic effective signal subset, the mechanism reference spectrum, and the single-machine correction vector. The coupling state analysis module is used to generate a residual-health coupling state matrix based on the equivalent residual package and health factors and to identify degenerate segments. The initial precursor transition graph construction module is used to construct the initial precursor transition graph based on the temporal adjacency relationship between degenerate segments, the continuity relationship of operating condition anchor points, the consistency relationship of residual direction, the maintenance of reversibility relationship, the multi-source support relationship, and the causal feasible region. The causal pruning and critical path identification module is used to perform causal direction identification, confusion factor detection, conditional independence test, causal strength coefficient calculation, causal edge confidence determination, path pruning, and extraction of dominant causal transfer sequence and optimal blocking point scheme under the same working condition anchor point. The downgraded causal chain reconstruction module is used to perform downgraded causal chain reconstruction based on the missing mode and generate a downgraded causal precursor transition map when the dynamic effective signal subset does not meet the normal prediction calculation conditions. The risk fusion assessment module is used to generate health index, fault risk score, risk status label and suggested work order; The intervention feedback self-correction module is used to update the causal edge confidence, causal strength coefficient, fault rule template weight, and optimal blocking point scheme score based on intervention feedback data after the suggested work order is executed. The system adopts an online hybrid deployment method and includes at least a field edge processing unit and a central analysis and processing unit. The field edge processing unit is used to perform data access, dynamic effective signal filtering, missing pattern identification, downgraded causal chain reconstruction, and conservative update calculation. The central analysis and processing unit is used to perform device group-level mechanism reference spectrum update, causal pruning and critical path identification, intervention feedback self-correction, and risk fusion assessment.

Images