A fault diagnosis method and system for a centrifugal compressor

Abstract

The application discloses a kind of fault diagnosis method and system of centrifugal compressor, comprising the following steps: S1: obtaining the equipment level information of target centrifugal compressor;S2: obtain the mechanical monitoring data and process monitoring data of the target centrifugal compressor;S3: construct multi-channel vibration event;S4: obtain the waveform spectrum feature and shaft center trajectory feature corresponding to the representative moment;S5: obtain feature label;S6: generate process-vibration coupling evidence;S7: homologous deduplication is carried out to feature label and process-vibration coupling evidence;S8: output corresponding fault evidence and disposal suggestion.The application can reduce the interference of time irrelevant or same source evidence on diagnostic result, improve the distinguishing ability between surge, rotating stall, oil film whirling, airflow disturbance and mechanical failure, and make fault diagnosis result and its formation basis easy to verify.

Description

Technical Field

[0001] This invention relates to the field of compressor technology, and in particular to a fault diagnosis method and system for centrifugal compressors. Background Technology

[0002] Centrifugal compressors are widely used in the petrochemical, energy, and process industries. Their operating status is typically monitored using data such as shaft vibration, shaft displacement, bearing temperature, lubrication parameters, inlet flow rate, inlet pressure, inlet temperature, and anti-surge valve opening. Existing fault diagnosis methods mostly rely on a single vibration threshold, local spectrum characteristics, shaft trajectory, or process parameters to determine equipment faults. Some methods also use weighted analysis of multiple monitoring parameters.

[0003] However, the sampling periods, time sources, and data quality of different monitoring data vary, and directly combining parameters can easily lead to temporally unrelated anomalies being mistaken for evidence of the same fault. Surge, rotational stall, oil film turbulence, airflow disturbance, and mechanical faults may exhibit similar low-frequency vibration or vibration fluctuation characteristics, making accurate differentiation difficult using only a single type of signal. Furthermore, the same abnormal event may generate duplicate evidence at multiple measurement points, across multiple time windows, and in multiple feature dimensions; directly accumulating scores can easily affect the diagnostic results.

[0004] Furthermore, existing methods have difficulty distinguishing between normal surge valve regulation, fluid disturbance, and mechanical failure based on process parameters and the order of mechanical vibration occurrence. Therefore, it is necessary to propose a centrifugal compressor fault diagnosis method and system that can perform time correlation, evidence deduplication, and stratified diagnosis on multi-source monitoring data. Summary of the Invention

[0005] The purpose of this invention is to provide a fault diagnosis method and system for centrifugal compressors to solve the above-mentioned problems.

[0006] To achieve the above objectives, the following technical solution is adopted: A fault diagnosis method for a centrifugal compressor includes the following steps: S1: Obtain the equipment level information of the target centrifugal compressor, determine the measuring points, bearings and rotors corresponding to the target diagnostic object based on the equipment level information, and establish the membership relationship between the measuring points, bearings, rotors and the target centrifugal compressor; S2: Obtain the mechanical monitoring data and process monitoring data of the target centrifugal compressor, convert the timestamps of different data sources into a unified event time, and determine the validity of each monitoring data based on time quality and data quality; S3: Identify abnormal shaft vibration events from the shaft vibration trend sequence; construct multi-channel vibration events based on the bearing position, detection direction, and time relationship between abnormal shaft vibration events corresponding to different shaft vibration channels; identify process events from the process monitoring data respectively; S4: Select a representative moment for the identified shaft vibration abnormal event, and obtain the waveform spectrum features and shaft center trajectory features corresponding to the representative moment; S5: Convert trend features, multi-channel vibration events, process events, waveform spectrum features, and shaft trajectory features into feature labels with fault physical meaning, and associate feature labels generated from the same abnormal event with the same original event identifier; S6: Based on the preset centrifugal compressor fault rules, calculate the start time difference, event interval overlap ratio, peak time difference and recovery order between process events and multi-channel vibration events, and generate process-vibration coupling evidence based on the calculation results and the directional timing conditions configured according to the corresponding fault rules. S7: Deduplicatize the feature tags and process-vibration coupling evidence from the same source, and aggregate the evidence according to the hierarchical order of measuring point, bearing, rotor and target centrifugal compressor; S8: Filter valid candidate faults based on the key data, necessary conditions, directional timing conditions and exclusionary counter-evidence conditions required for each candidate fault. Score the valid candidate faults based on the supporting conditions and ordinary counter-evidence conditions. Determine the main fault and candidate faults based on the score results and the exclusion relationship between the candidate faults. Output the corresponding fault evidence and handling suggestions.

[0007] Furthermore, in S2, the mechanical monitoring data includes at least shaft vibration data from multiple shaft vibration channels, and includes at least one of shaft displacement data, bearing temperature data, lubricating oil temperature data, and rotational speed data; the process monitoring data includes at least one of inlet flow rate, inlet pressure, inlet temperature, and anti-surge valve opening; key data and auxiliary data are configured for different candidate faults, and when key data is missing or invalid, the corresponding candidate fault is excluded from the valid candidate set; when auxiliary data is missing or invalid, the supporting conditions corresponding to the auxiliary data are stopped or the evidence level of the corresponding candidate fault is reduced.

[0008] Furthermore, in S2, the timestamps from different data sources are converted into a unified event time, including the following steps: The system acquires the data source timestamp, acquisition timestamp, sampling period, engineering unit, and quality code of various monitoring data. It then corrects the data source timestamp according to the unified clock on the equipment side and records the clock deviation of the data source clock relative to the unified clock on the equipment side. The original sampling period of each monitoring data is retained, and the shaft vibration trend data and process trend data are resampled to a unified time grid. For continuous data, previous value hold or interpolation is used, and for discrete state data, previous value hold is used. When the continuous absence of measurement exceeds the absence tolerance of the corresponding data source, the quality code is invalid, the clock deviation exceeds the preset time correction tolerance, or the engineering unit cannot reach a consensus, the corresponding monitoring data will be marked as unavailable, and the use of unavailable monitoring data to meet the necessary conditions for candidate faults will be prohibited.

[0009] Furthermore, in S3, multi-channel vibration events are constructed, including the following steps: Identify abnormal shaft vibration events from each shaft vibration channel, and determine the start time, end time, and duration of each abnormal shaft vibration event; When the maximum difference in the start time of each abnormal shaft vibration event within the same channel group does not exceed the synchronization tolerance, and the number of abnormal channels reaches a preset number or the proportion of abnormal channels to effective channels reaches a preset proportion, the corresponding abnormal shaft vibration events are merged into a multi-channel vibration event. The channel group includes at least one of the following: two orthogonal direction shaft vibration channel groups for the same bearing, a shaft vibration channel group for the same rotor, and a whole-machine shaft vibration channel group for the target centrifugal compressor. The shaft vibration channel group for the same rotor includes two orthogonal direction shaft vibration channels corresponding to the first bearing position and the second bearing position, respectively. The first bearing position and the second bearing position are the connecting end bearing position and the non-connecting end bearing position of the rotor, respectively. When both orthogonal direction shaft vibration channels are effective and both exhibit abnormal shaft vibration events within the synchronization tolerance, a four-channel synchronous event is generated.

[0010] Furthermore, in S4, a representative moment is selected for the identified shaft vibration anomaly event, including the following steps: The corresponding time selection rule is selected according to the event type of shaft vibration abnormal event. Specifically, for abrupt type shaft vibration abnormal events, the time when the sudden change occurs is selected; for continuous fluctuation type shaft vibration abnormal events, the time when the fluctuation is the greatest is selected; and for stable high-level type shaft vibration abnormal events, the time within the stable abnormal range is selected. The original waveform data of the corresponding shaft vibration channel is retrieved according to the representative moment, and the waveform spectrum characteristics are calculated based on the original waveform data; synchronous original waveform data of the shaft vibration channels in two orthogonal directions of the same bearing are retrieved, and the shaft center trajectory is generated and the shaft center trajectory characteristics are extracted based on the synchronous original waveform data of the two orthogonal directions; or the waveform spectrum characteristics and shaft center trajectory characteristics generated by the external monitoring system according to the representative moment are received; wherein, the waveform spectrum characteristics include at least one of order spectrum characteristics, low frequency characteristics, phase characteristics, clipping characteristics, glitch characteristics, peak-valley asymmetry characteristics and baseline drift characteristics, and the shaft center trajectory characteristics include at least one of precession direction, trajectory repeatability, trajectory axis length ratio, trajectory area and trajectory shape; wherein, the fundamental frequency of the order spectrum is determined based on the rotational speed data verified by data quality.

[0011] Furthermore, in S5, trend features, multi-channel vibration events, process events, waveform spectrum features, and shaft trajectory features are encapsulated into standard feature samples with measurement point identifiers, bearing identifiers, rotor identifiers, event times, original event identifiers, quality codes, and feature values, respectively. The standard feature samples are then converted into corresponding feature labels according to preset thresholds and feature combination conditions. For multiple feature labels generated from the same abnormal event, the same source deduplication is performed at the measurement point layer according to the original event identifier, and when aggregating step by step to the bearing layer, rotor layer and target centrifugal compressor layer, the same abnormal event is avoided from being counted repeatedly according to the original event identifier; For the same abnormal event, feature labels generated in different trend time windows are retained, and the feature labels are merged into a set of evidence in the same fault rule.

[0012] Furthermore, in S6, generating process-vibration coupling evidence includes the following steps: Identify the corresponding process events from the trend data of inlet flow rate, inlet pressure, inlet temperature, and anti-surge valve opening, respectively; Based on the order of occurrence and duration of process events, multiple related process events are combined into a composite process event, and a single process event or the composite process event is identified as a process event to be associated. Calculate the start time difference, event interval overlap ratio, peak time difference, and recovery order of the process event to be associated relative to the multi-channel vibration event; When the initial time difference is within the directional allowable lag interval set by the corresponding fault rule, and the event interval overlap ratio, peak time difference, or recovery order meets the corresponding judgment conditions, the process-vibration coupling evidence is generated.

[0013] Furthermore, in S6, the centrifugal compressor fault rules include key data conditions, necessary conditions, supporting conditions, ordinary counter-evidence conditions, exclusionary counter-evidence conditions, and fault exclusion relationships. The necessary conditions are formed by combining feature tags and process-vibration coupling evidence according to AND logic, OR logic, logic that satisfies at least a preset number, or logic that satisfies at least a preset proportion. When the key data is valid, the necessary conditions are met, the directional timing conditions are met, and the exclusionary counter-evidence conditions are not hit, optional bonus points are determined according to the supporting conditions, and counter-evidence deduction points are determined according to the ordinary counter-evidence conditions. The candidate fault scores are obtained under the limitation of the optional bonus point cap and the maximum total score of the fault rules. The scores of different candidate faults are normalized. The main fault and candidate faults are determined according to the normalized scores, fault exclusion relationships, and preset fault priorities. The corresponding fault evidence is retained for faults that can coexist.

[0014] A fault diagnosis system for a centrifugal compressor, to implement the above-mentioned method, includes: The equipment analysis and data acquisition module is used to acquire equipment hierarchy information, mechanical monitoring data and process monitoring data of the target centrifugal compressor, and to establish the hierarchical relationship between measuring points, bearings, rotor and the target centrifugal compressor. The time processing and event construction module is used to convert timestamps from different data sources into unified event times, determine the validity of various monitoring data based on time quality and data quality, identify shaft vibration abnormal events from shaft vibration trend sequences and construct multi-channel vibration events, and identify process events from process monitoring data. The deep feature acquisition module is used to select a representative moment for the shaft vibration abnormal event and acquire the waveform spectrum features and shaft center trajectory features corresponding to the representative moment. The feature label generation module is used to convert trend features, multi-channel vibration events, process events, waveform spectrum features, and shaft center trajectory features into feature labels, and associate feature labels generated from the same abnormal event with the same original event identifier. The coupling and evidence aggregation module is used to generate process-vibration coupling evidence based on the start time difference, event interval overlap ratio, peak time difference and recovery order between process events and multi-channel vibration events. The module performs deduplication of the feature tags and process-vibration coupling evidence, and aggregates the evidence according to the hierarchical order of measuring point, bearing, rotor and target centrifugal compressor. The rule diagnosis and result output module is used to screen valid candidate faults based on key data, necessary conditions, directional timing conditions and exclusionary counter-evidence conditions, score valid candidate faults based on supporting conditions and ordinary counter-evidence conditions, determine the primary fault and candidate faults based on the scoring results and fault exclusion relationships, and output corresponding fault evidence and handling suggestions.

[0015] Furthermore, it also includes a rule calibration module and a rule configuration library; the rule calibration module is used to perform offline calibration of the judgment threshold, necessary conditions, supporting conditions, counter-evidence conditions and scoring weights in the initial fault rules established by fault mechanisms and expert experience based on historical fault samples with fault truth value labels, and writes the reviewed fault rules into the rule configuration library; the rule configuration library is used to configure corresponding fault rules according to the centrifugal compressor model, historical baseline of measuring points and current operating conditions, and the rule diagnosis and result output module performs online diagnosis or offline diagnosis according to the reviewed fault rules in the rule configuration library.

[0016] By adopting the above solution, the beneficial effects of the present invention are: This invention converts mechanical and process monitoring data of centrifugal compressors to a unified event time. It constructs multi-channel vibration events based on shaft vibration anomalies and forms directional coupled evidence by considering the start time difference, event interval overlap ratio, peak time difference, and recovery order between process events and multi-channel vibration events. Simultaneously, it extracts waveform spectrum features and shaft trajectory features from representative moments of the anomaly events, converting features from different sources into traceable feature labels. After deduplication based on the same source and aggregation at the measurement point, bearing, rotor, and equipment levels, it further combines necessary conditions, supporting conditions, and counter-evidence conditions to screen and score candidate faults. This reduces the interference of temporally irrelevant or identically sourced evidence on diagnostic results, improves the ability to distinguish between surge, rotating stall, oil film eddy, airflow disturbance, and mechanical faults, and makes the fault diagnosis results and their basis for verification easier. Attached Figure Description

[0017] Figure 1 This is a flowchart of the method of the present invention; Figure 2 This is a system architecture diagram of the present invention. Detailed Implementation

[0018] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments.

[0019] Reference Figures 1 to 2 As shown, in one embodiment, the present invention provides a fault diagnosis method for a centrifugal compressor, comprising the following steps: S1: Obtain the equipment level information of the target centrifugal compressor, determine the measuring points, bearings and rotors corresponding to the target diagnostic object based on the equipment level information, and establish the membership relationship between the measuring points, bearings, rotors and the target centrifugal compressor.

[0020] In step S1, the target diagnostic object can be a rotor or the entire centrifugal compressor. The system parses the equipment tree according to the equipment identifier, determines the bearings, rotors and equipment corresponding to each shaft vibration measuring point, shaft displacement measuring point, temperature measuring point and process measuring point, and establishes the membership relationship between measuring point, bearing, rotor and centrifugal compressor.

[0021] For centrifugal compressors with multiple rotors, a diagnostic context can be established for each rotor. First, rotor-level diagnostic results are generated, and then evidence aggregation is performed at the whole machine level. The bearing position closer to the connection side is defined as the connection-end bearing position, and the bearing position farther from the connection side is defined as the non-connection-end bearing position. Two orthogonal shaft vibration channels are set for the connection-end bearing position and the non-connection-end bearing position, respectively. The connection-end X channel, connection-end Y channel, non-connection-end X channel and non-connection-end Y channel constitute the four-channel event group of the target rotor.

[0022] The equipment-level information is also used to distinguish between key data and auxiliary data. Different candidate faults can be configured with the required key data and auxiliary data respectively. When the key data corresponding to a candidate fault is missing or invalid, other weak evidence will no longer be used to supplement the candidate fault, and the candidate fault will not enter the valid candidate set. When the auxiliary data is missing or invalid, only the corresponding supporting conditions will be stopped, or the evidence level of the candidate fault will be reduced.

[0023] S2: Obtain the mechanical monitoring data and process monitoring data of the target centrifugal compressor, convert the timestamps of different data sources into a unified event time, and determine the validity of each monitoring data based on time quality and data quality.

[0024] In S2, the mechanical monitoring data includes at least shaft vibration data from multiple shaft vibration channels, and includes at least one of shaft displacement data, bearing temperature data, lubricating oil temperature data, and rotational speed data; the process monitoring data includes at least one of inlet flow rate, inlet pressure, inlet temperature, and anti-surge valve opening; key data and auxiliary data are configured for different candidate faults, and when key data is missing or invalid, the corresponding candidate fault is excluded from the valid candidate set; when auxiliary data is missing or invalid, the supporting conditions corresponding to the auxiliary data are stopped or the evidence level of the corresponding candidate fault is reduced.

[0025] In S2, converting timestamps from different data sources into a unified event time includes the following steps: The system acquires the data source timestamp, acquisition timestamp, sampling period, engineering unit, and quality code of various monitoring data. It then corrects the data source timestamp according to the unified clock on the equipment side and records the clock deviation of the data source clock relative to the unified clock on the equipment side. The original sampling period of each monitoring data is retained, and the shaft vibration trend data and process trend data are resampled to a unified time grid. For continuous data, previous value hold or interpolation is used, and for discrete state data, previous value hold is used. When the continuous absence of measurement exceeds the absence tolerance of the corresponding data source, the quality code is invalid, the clock deviation exceeds the preset time correction tolerance, or the engineering unit cannot reach a consensus, the corresponding monitoring data will be marked as unavailable, and the use of unavailable monitoring data to meet the necessary conditions for candidate faults will be prohibited.

[0026] In this step, the mechanical monitoring data and process monitoring data of the target centrifugal compressor are mainly acquired, and the time of different data sources is unified and the data quality is judged.

[0027] The mechanical monitoring data includes at least shaft vibration data from multiple shaft vibration channels, and may also include shaft displacement data, bearing temperature data, lubricating oil temperature data, and rotational speed data; the process monitoring data may include inlet flow rate, inlet pressure, inlet temperature, and anti-surge valve opening. The system selects the corresponding data type according to different fault rules.

[0028] The system establishes multiple trend time windows around the diagnostic time. As one implementation method, three time windows can be established: 1 day, 3 days, and 30 days. The shorter time window is used to identify transient fluctuations and sudden changes, while the longer time window is used to identify sustained high levels, long-term changes, and recurring anomalies. The number and length of the time windows can be configured according to the equipment's operating cycle.

[0029] Each monitoring data point carries a data source timestamp, acquisition timestamp, measurement point identifier, engineering unit, and quality code. The system uses the unified clock on the equipment side as a reference to correct the timestamps of each data source and records the clock deviation of the data source clock relative to the unified clock on the equipment side. When the data source does not provide a valid data source timestamp, the acquisition timestamp can be used as the event time, but the time quality of the data is marked as downgraded so that it cannot be used to form high-level directional coupling evidence.

[0030] In addition, the system retains the original sampling period of each data source and resamples shaft vibration trends and process trends to a unified time grid according to diagnostic needs. For example, the original sampling period for shaft vibration trends can be 1 second, the sampling period for inlet flow rate, inlet pressure, and anti-surge valve opening can be 1 to 5 seconds, and the sampling period for temperature data can be 10 to 60 seconds. A 1-second unified time grid can be used for rapidly changing trend data, continuous data with sampling periods longer than the unified time grid can use previous value hold or linear interpolation, and discrete data such as anti-surge valve status uses previous value hold. The original waveform is not resampled at low frequencies; instead, a correspondence is established between the waveform trigger time and trend events.

[0031] When the time interval between adjacent valid data exceeds the missing test tolerance of the corresponding data source, the system will not perform interpolation across the missing test interval; when the quality code is invalid, the clock deviation exceeds the preset time correction tolerance, or the engineering unit cannot reach a consensus, the corresponding data will be marked as unusable, and the data marked as unusable must not meet the necessary conditions for the candidate fault.

[0032] A general judgment threshold can be written into the rule configuration library as an initial threshold and corrected based on the centrifugal compressor model, historical baseline of the measuring points, and current operating conditions. The current operating conditions can include speed, load, inlet medium, inlet temperature, and environmental conditions. Therefore, different judgment thresholds can be used for the same type of feature under different units or different operating conditions.

[0033] S3: Identify abnormal shaft vibration events from the shaft vibration trend sequence; construct multi-channel vibration events based on the bearing position, detection direction, and time relationship between abnormal shaft vibration events corresponding to different shaft vibration channels; identify process events from the process monitoring data.

[0034] In S3, constructing multi-channel vibration events includes the following steps: Identify abnormal shaft vibration events from each shaft vibration channel, and determine the start time, end time, and duration of each abnormal shaft vibration event; When the maximum difference in the start time of each abnormal shaft vibration event within the same channel group does not exceed the synchronization tolerance, and the number of abnormal channels reaches a preset number or the proportion of abnormal channels to effective channels reaches a preset proportion, the corresponding abnormal shaft vibration events are merged into a multi-channel vibration event. The channel group includes at least one of the following: two orthogonal direction shaft vibration channel groups for the same bearing, a shaft vibration channel group for the same rotor, and a whole-machine shaft vibration channel group for the target centrifugal compressor. The shaft vibration channel group for the same rotor includes two orthogonal direction shaft vibration channels corresponding to the first bearing position and the second bearing position, respectively. The first bearing position and the second bearing position are the connecting end bearing position and the non-connecting end bearing position of the rotor, respectively. When both orthogonal direction shaft vibration channels are effective and both exhibit abnormal shaft vibration events within the synchronization tolerance, a four-channel synchronous event is generated.

[0035] In this step, trend features are mainly extracted from the shaft vibration trend sequence and abnormal shaft vibration events are identified, while process events are identified from process monitoring data.

[0036] The trend features can include the current value, window mean, fluctuation score, number of upward segments, jump amplitude, number of abnormal changes, peaks, pullbacks, and change point types. Based on these trend features, the system converts the continuous trend curve into a trend snapshot that can be used for event identification.

[0037] The fluctuation score is used to characterize the degree of change in monitoring data within a preset sliding window. The fluctuation score can be determined based on one or more of the following: standard deviation, range, rate of change between adjacent sampling points, and number of abnormal changes. For example, the ratio of the standard deviation to the window mean can be used as the fluctuation score, or the range, rate of change, and number of abnormal changes can be normalized separately and then weighted and summed. The indicators used and their weights can be configured according to the type of measurement point and the operating conditions.

[0038] Different identification criteria can be used for the start and end times of different types of abnormal events. For jump events, the start time can be determined when the change in adjacent sampling points first reaches the jump threshold, and the end time can be determined when the monitored value re-enters the stable range. For continuous fluctuation events, the start time can be determined when a preset number of consecutive sampling points reach the fluctuation condition, and the end time can be determined when a preset number of consecutive sampling points return to normal. For stable high-level events, the interval in which the monitored value continuously exceeds the high-value threshold can be determined as the event duration interval. By determining the start time, end time, and duration interval of abnormal events for each channel, a foundation is provided for subsequent multi-channel event merging.

[0039] In addition, the system establishes two orthogonal direction shaft vibration channel groups for the same bearing, a shaft vibration channel group for the same rotor, and a whole machine shaft vibration channel group. For multiple shaft vibration abnormal events within the same channel group, when the maximum difference between the start times of each shaft vibration abnormal event does not exceed the synchronization tolerance, and the number of abnormal channels reaches a preset number or the proportion of abnormal channels to effective channels reaches a preset proportion, the corresponding shaft vibration abnormal events are merged into multi-channel vibration events.

[0040] As an example of parameters, the synchronization tolerance can be set to 3 seconds, the minimum number of abnormal channels can be set to 2, and the minimum percentage of abnormal channels can be set to 0.5. When at least two shaft vibration channels or no less than 50% of all effective shaft vibration channels experience corresponding abnormalities within 3 seconds, a multi-channel synchronization event is formed.

[0041] Four-channel synchronization events are recorded separately from general multi-channel synchronization events. A four-channel synchronization event is generated only when all four channels (connected X channel, connected Y channel, non-connected X channel, and non-connected Y channel) are valid and shaft vibration abnormal events occur within the synchronization tolerance. Only when the minimum number of abnormal channels or the minimum percentage of abnormal channels is reached, a multi-channel synchronization event is generated, and it is not equivalent to a four-channel synchronization event.

[0042] If only one shaft vibration channel exhibits a sudden jump, constant value, clipping, over-range, or abnormal quality code, while the other orthogonal direction channel of the same bearing, other shaft vibration channels of the same rotor, and corresponding process quantities do not change accordingly, then this constitutes isolated single-channel anomaly evidence. This evidence can serve as ordinary or exclusionary counter-evidence of mechanical failure and as supporting evidence of sensor anomaly.

[0043] The system also identifies process events from the trend data of inlet flow, inlet pressure, inlet temperature and anti-surge valve opening. Process events can include inlet flow decrease events, inlet pressure low-frequency fluctuation events, inlet temperature change events, anti-surge valve opening increase events and anti-surge valve repeated operation events.

[0044] S4: Select a representative moment for the identified shaft vibration abnormal event, and obtain the waveform spectrum features and shaft center trajectory features corresponding to the representative moment.

[0045] In S4, a representative moment is selected for the identified shaft vibration anomaly event, including the following steps: The corresponding time selection rule is selected according to the event type of shaft vibration abnormal event. Specifically, for abrupt type shaft vibration abnormal events, the time when the sudden change occurs is selected; for continuous fluctuation type shaft vibration abnormal events, the time when the fluctuation is the greatest is selected; and for stable high-level type shaft vibration abnormal events, the time within the stable abnormal range is selected. The original waveform data of the corresponding shaft vibration channel is retrieved according to the representative moment, and the waveform spectrum characteristics are calculated based on the original waveform data; synchronous original waveform data of the shaft vibration channels in two orthogonal directions of the same bearing are retrieved, and the shaft center trajectory is generated and the shaft center trajectory characteristics are extracted based on the synchronous original waveform data of the two orthogonal directions; or the waveform spectrum characteristics and shaft center trajectory characteristics generated by the external monitoring system according to the representative moment are received; wherein, the waveform spectrum characteristics include at least one of order spectrum characteristics, low frequency characteristics, phase characteristics, clipping characteristics, glitch characteristics, peak-valley asymmetry characteristics and baseline drift characteristics, and the shaft center trajectory characteristics include at least one of precession direction, trajectory repeatability, trajectory axis length ratio, trajectory area and trajectory shape; wherein, the fundamental frequency of the order spectrum is determined based on the rotational speed data verified by data quality.

[0046] In this step, representative moments are selected for the identified shaft vibration anomaly events, and waveform spectrum features and shaft center trajectory features are obtained around the representative moments.

[0047] Specifically, the system can select at least one representative moment from each shaft vibration anomaly event, and different time selection rules are used for different types of anomalies. For abrupt shaft vibration anomalies, the moment of abrupt change is selected; for continuous fluctuation shaft vibration anomalies, the moment with the largest fluctuation fraction is selected; for stable high-level shaft vibration anomalies, the moment within the stable anomaly interval is selected. Representative moments are selected for multiple anomalies, so that the deep features that recur at adjacent moments can form repeatable evidence, rather than making a judgment based solely on a single anomaly.

[0048] When the original waveform data is available, the system retrieves the original waveform data of the corresponding shaft vibration channel based on representative moments, and extracts features such as order spectrum, low-frequency components, phase, clipping, glitches, peak-valley asymmetry, and baseline drift from the original waveform data. The fundamental frequency of the order spectrum is determined based on the rotational speed data that has been quality verified, and the peak value of the maximum vibration spectrum that has not been confirmed is not directly used as the 1X frequency.

[0049] Waveform spectral characteristics can include the dominance of 1X, the amplitude ratio of 2X to 1X, the amplitude in the 0.38X to 0.48X frequency band, low-frequency components greater than 0.5X, ultra-low-frequency components from 1Hz to 30Hz, clipping ratio, glitch ratio, and peak-valley asymmetry ratio. As an example of parameters, a 1X dominance ratio of 0.7 can be considered a 1X dominance feature, a 2X to 1X amplitude ratio of 0.5 can be considered a 2X active feature, amplitudes reaching the corresponding thresholds in the 0.38X to 0.48X frequency band can be considered low-frequency characteristics of oil film whirl, a clipping ratio of 0.05 can be considered a clipping feature, and a peak-valley asymmetry ratio of 1.3 can be considered a peak-valley asymmetry feature.

[0050] The shaft center trajectory is generated from the synchronous raw waveform data of two orthogonal shaft vibration channels of the same bearing. The system extracts the precession direction, trajectory repeatability, trajectory axis length ratio, trajectory area, and trajectory morphology from the shaft center trajectory. As a parameter example, a trajectory repeatability score of 0.72 can be considered high repeatability, a score below 0.45 low repeatability, an axis length ratio of 2.2 elongated trajectory, and a trajectory area of ​​1300 square micrometers large area. The trajectory morphology can include elliptical, elongated, honeycomb, and other irregular shapes.

[0051] When the external condition monitoring system has completed waveform spectrum analysis or shaft center trajectory analysis, this system can directly receive the feature results generated by the external condition monitoring system. Regardless of whether the features are calculated by this system or provided by the external condition monitoring system, they will be processed in the same standard feature sample format.

[0052] S5: Convert trend features, multi-channel vibration events, process events, waveform spectrum features, and shaft trajectory features into feature labels with fault physical meaning, and associate feature labels generated from the same abnormal event with the same original event identifier.

[0053] In S5, trend features, multi-channel vibration events, process events, waveform spectrum features, and shaft trajectory features are encapsulated into standard feature samples with measurement point identifiers, bearing identifiers, rotor identifiers, event times, original event identifiers, quality codes, and feature values, respectively. The standard feature samples are then converted into corresponding feature labels according to preset thresholds and feature combination conditions. For multiple feature labels generated from the same abnormal event, the same source deduplication is performed at the measurement point layer according to the original event identifier, and when aggregating step by step to the bearing layer, rotor layer and target centrifugal compressor layer, the same abnormal event is avoided from being counted repeatedly according to the original event identifier; For the same abnormal event, feature labels generated in different trend time windows are retained, and the feature labels are merged into a set of evidence in the same fault rule.

[0054] In this step, each standard feature sample records at least the measuring point identifier, bearing identifier, rotor identifier, equipment identifier, event time, original event identifier, time window, quality code, engineering unit, feature type, and feature value. Trend features, multi-channel features, waveform spectrum features, and shaft trajectory features obtained from the same abnormal event use a common original event identifier, thereby enabling tracing back to the corresponding monitoring data and event time.

[0055] The system converts standard feature samples into feature labels with physical meaning of faults based on preset thresholds and feature combination conditions. For example, when a certain shaft vibration channel is continuously higher than the high value threshold and the fluctuation score is lower than the corresponding threshold within a long-term time window, a stable high-level vibration label is generated; when the amplitude of the 0.38X to 0.48X frequency band in the waveform spectrum reaches the corresponding threshold, an oil film whirl low-frequency feature label is generated; when multiple shaft vibration channels exhibit rapid fluctuations within the synchronization tolerance, a multi-channel rapid fluctuation label is generated; when the inlet flow rate decrease event and the multi-channel vibration event meet the directional timing condition, an inlet flow rate decrease and vibration coupling label is generated. In this embodiment, the feature labels are used to represent the physical characteristics presented by the current monitoring data, and the confirmed fault types in historical fault cases are used as fault truth labels for offline calibration of judgment thresholds, rule conditions, and scoring weights.

[0056] S6: Based on the preset centrifugal compressor fault rules, calculate the start time difference, event interval overlap ratio, peak time difference and recovery order between process events and multi-channel vibration events, and generate process-vibration coupling evidence based on the calculation results and the directional timing conditions configured according to the corresponding fault rules.

[0057] In S6, the centrifugal compressor fault rules include key data conditions, necessary conditions, supporting conditions, ordinary counter-evidence conditions, exclusionary counter-evidence conditions, and fault exclusion relationships. The necessary conditions are formed by combining feature tags and process-vibration coupling evidence according to AND logic, OR logic, logic that satisfies at least a preset number, or logic that satisfies at least a preset proportion. When the key data is valid, the necessary conditions are met, the directional timing conditions are met, and the exclusionary counter-evidence conditions are not hit, optional bonus points are determined according to the supporting conditions, and counter-evidence deduction points are determined according to the ordinary counter-evidence conditions. The candidate fault scores are obtained under the limit of the optional bonus points and the maximum total score of the fault rules. The scores of different candidate faults are normalized. The main fault and candidate faults are determined according to the normalized scores, fault exclusion relationships, and preset fault priorities. The corresponding fault evidence is retained for faults that can coexist.

[0058] In S6, generating evidence of process-vibration coupling includes the following steps: Identify the corresponding process events from the trend data of inlet flow rate, inlet pressure, inlet temperature, and anti-surge valve opening, respectively; Based on the order of occurrence and duration of process events, multiple related process events are combined into a composite process event, and a single process event or the composite process event is identified as a process event to be associated. Calculate the start time difference, event interval overlap ratio, peak time difference, and recovery order of the process event to be associated relative to the multi-channel vibration event; When the initial time difference is within the directional allowable lag interval set by the corresponding fault rule, and the event interval overlap ratio, peak time difference, or recovery order meets the corresponding judgment conditions, the process-vibration coupling evidence is generated.

[0059] In this step, the system first identifies various individual process events, and then combines multiple related process events into composite process events based on their occurrence order and duration. For example, a decrease in inlet flow rate, low-frequency fluctuations in inlet pressure, and repeated operation of the anti-surge valve can be combined into a composite process event. Both single and composite process events can be compared with multi-channel vibration events as process events to be correlated.

[0060] Let the start time of the process event to be correlated be the process start time, and the start time of the multi-channel vibration event be the vibration start time. The initial time difference between the two is the difference between the process start time and the vibration start time. When the initial time difference is negative, it means that the process event occurs before the vibration event. When the initial time difference is positive, it means that the vibration event occurs before the process event. Different fault rules are configured with directional allowable lag intervals to distinguish different situations such as the inlet flow rate decreasing first, the anti-surge valve acting first, and the vibration abnormality occurring first.

[0061] The overlap ratio of event intervals can be determined based on the ratio of the overlap duration between the process event to be associated and the multi-channel vibration event to the duration of the shorter event, or it can be determined based on the ratio of the overlap duration to the union duration of the two event intervals. The specific calculation method should be consistent within the same rule configuration. The peak time difference is the difference between the peak time of the abnormal process parameter and the peak time of the abnormal vibration. The recovery order is used to indicate the order in which the process quantity recovers to stability and the vibration recovers to stability.

[0062] When the initial time difference is within the directional allowable lag range set by the corresponding fault rule, and the event interval overlap ratio, peak time difference, or recovery order meets the corresponding conditions, process-vibration coupling evidence is generated. Process events and vibration events that do not meet the synchronization condition or directional lag condition in time shall not be used as supporting evidence for the same fault event for joint scoring.

[0063] For anti-surge valve operation, the system can also read valve control commands or target openings. When the anti-surge valve changes monotonically and at a limited speed according to the control command, and the inlet flow and intake pressure stabilize again within the allowable recovery time, and no multi-channel vibration event meets the channel number requirements, the valve operation is identified as normal regulation, forming counter-evidence for surge candidates. When the anti-surge valve exhibits rapid or repetitive operation beyond the control dead zone, and forms a directional coupling with a decrease in inlet flow, low-frequency fluctuations in intake pressure, and repetitive multi-channel vibration events, it is used as supporting evidence for surge or surge precursors. When valve control commands cannot be obtained, the weight of valve position change evidence is reduced, and valve position change alone does not satisfy the necessary conditions for surge.

[0064] S7: Deduplicatize the feature labels and process-vibration coupling evidence from the same source, and aggregate the evidence according to the hierarchical order of measurement point, bearing, rotor and target centrifugal compressor.

[0065] In this step, different types of labels are statistically analyzed according to different sample units. Waveform spectrum labels can be statistically analyzed based on a single measurement point and a single representative moment; shaft center trajectory labels can be statistically analyzed based on a single bearing and a single representative moment; trend labels can be statistically analyzed based on a single measurement point and a single time window; multi-channel labels can be statistically analyzed based on a single channel group and a single event; and process-vibration coupling evidence can be statistically analyzed based on the correlation between a single process event to be associated and a single multi-channel vibration event.

[0066] The same original event can generate multiple feature labels, but each feature label retains the common original event identifier. The system first performs deduplication based on the original event identifier at the measurement point layer, and then aggregates it level by level to the bearing layer, rotor layer and whole machine layer. When aggregating to the next level, it continues to judge whether each piece of evidence comes from the same abnormal event based on the original event identifier, so as to avoid the same abnormal event being repeatedly scored due to the formation of multiple feature labels.

[0067] The same abnormal event may also appear in multiple trend time windows at the same time. The system retains the feature labels corresponding to different time windows to indicate the state of the abnormality at different time scales. However, in the same fault rule, cross-window labels from the same abnormal event are merged into a set of evidence. Multiple independent abnormal events can form event-level evidence chains respectively, and are aggregated in the bearing layer, rotor layer and whole machine layer according to their repetition frequency, duration and spatial distribution.

[0068] The aggregation results can include label counts, the proportion of labels in valid samples, the presence of flags for the corresponding labels, and the number of device-level coupling events. Flag-type evidence can correspond to a fixed score, count-type evidence can be accumulated in segments or with a capped value, and proportion-type evidence can be mapped to a corresponding score based on its interval.

[0069] S8: Filter valid candidate faults based on the key data, necessary conditions, directional timing conditions and exclusionary counter-evidence conditions required for each candidate fault. Score the valid candidate faults based on the supporting conditions and ordinary counter-evidence conditions. Determine the main fault and candidate faults based on the score results and the exclusion relationship between the candidate faults. Output the corresponding fault evidence and handling suggestions.

[0070] In this step, the centrifugal compressor fault rules may include key data conditions, necessary conditions, supporting conditions, ordinary counter-evidence conditions, exclusionary counter-evidence conditions, scoring parameters, fault exclusion relationships, and fault priorities. Necessary conditions may be combined using AND logic, OR logic, logic that satisfies at least a preset number, or logic that satisfies at least a preset proportion. Different candidate faults use different combinations of conditions, and it is not required that all necessary labels be judged in the same way.

[0071] The system first checks whether the key data required for a candidate fault is valid, then determines whether the necessary conditions and directional timing conditions are met, and checks whether the exclusionary proof condition is met. If the key data is invalid, the necessary conditions are not met, the directional timing conditions are not met, or the exclusionary proof condition is met, the candidate fault is not included in the valid candidate set. For candidate faults that are included in the valid candidate set, optional bonus points are determined based on the supporting conditions, and proof-of-contrast deduction points are determined based on the ordinary proof-of-contrast conditions.

[0072] In this invention, for any candidate fault rule r, let the necessary condition of the r-th fault rule be satisfied as . The value is 1 if the condition is met, and 0 if the condition is not met; the basic values ​​are... The total optional bonus points are: The maximum optional bonus points is: The total deduction for ordinary counter-evidence conditions is: The maximum total score is The final score is The final score can then be determined using the following formula: , when When the score is 0, the final score is 0, and the corresponding fault does not enter the valid candidate set. As an example of parameters, the base score can be set between 0.45 and 0.65, the maximum total score can be set between 0.85 and 0.95, and the optional bonus score cannot exceed the difference between the maximum total score and the base score. The above parameters can be calibrated offline based on historical cases.

[0073] When different fault rules have different maximum total scores, a normalized score can be formed by the ratio of the final score to the corresponding maximum total score, and then candidates can be ranked based on the normalized score. As one implementation method, a normalized score of 0.82 is considered a high level of evidence, 0.65 or lower is considered a medium level of evidence, and lower than 0.65 is considered a low level of evidence. The level of evidence indicates the relative sufficiency of the evidence for the current rule, and does not directly represent the statistically validated diagnostic accuracy.

[0074] When multiple candidate faults simultaneously meet the output conditions, the system first eliminates conflicting candidates based on fault exclusion relationships, then determines the primary fault by combining normalized scores and fault priorities, and the remaining faults that were not eliminated are treated as candidate faults. For compound faults that can coexist, their scores and chains of evidence are retained separately.

[0075] The diagnostic results include the main fault, candidate faults, normalized score, level of evidence, supporting evidence, counter-evidence, source of evidence, operational recommendations, and maintenance recommendations. The system only outputs and saves the diagnostic results and treatment recommendations, without directly adjusting the compressor load, anti-surge valve, or executing shutdown control.

[0076] Furthermore, in another embodiment, a fault diagnosis system for a centrifugal compressor is also provided to implement the above-described method, including... The equipment analysis and data acquisition module is used to acquire equipment hierarchy information, mechanical monitoring data and process monitoring data of the target centrifugal compressor, and to establish the hierarchical relationship between measuring points, bearings, rotor and the target centrifugal compressor. The time processing and event construction module is used to convert timestamps from different data sources into unified event times, determine the validity of various monitoring data based on time quality and data quality, identify shaft vibration abnormal events from shaft vibration trend sequences and construct multi-channel vibration events, and identify process events from process monitoring data. The deep feature acquisition module is used to select a representative moment for the shaft vibration abnormal event and acquire the waveform spectrum features and shaft center trajectory features corresponding to the representative moment. The feature label generation module is used to convert trend features, multi-channel vibration events, process events, waveform spectrum features, and shaft center trajectory features into feature labels, and associate feature labels generated from the same abnormal event with the same original event identifier. The coupling and evidence aggregation module is used to generate process-vibration coupling evidence based on the start time difference, event interval overlap ratio, peak time difference and recovery order between process events and multi-channel vibration events. The module performs deduplication of the feature tags and process-vibration coupling evidence, and aggregates the evidence according to the hierarchical order of measuring point, bearing, rotor and target centrifugal compressor. The rule diagnosis and result output module is used to screen valid candidate faults based on key data, necessary conditions, directional timing conditions and exclusionary counter-evidence conditions, score valid candidate faults based on supporting conditions and ordinary counter-evidence conditions, determine the primary fault and candidate faults based on the scoring results and fault exclusion relationships, and output corresponding fault evidence and handling suggestions.

[0077] Simultaneously, it also includes a rule calibration module and a rule configuration library; the rule calibration module is used to perform offline calibration of the judgment threshold, necessary conditions, supporting conditions, counter-evidence conditions, and scoring weights in the initial fault rules established by fault mechanisms and expert experience based on historical fault samples with fault truth value labels, and writes the reviewed fault rules into the rule configuration library; the rule configuration library is used to configure corresponding fault rules according to the centrifugal compressor model, historical baseline of measuring points, and current operating conditions, and the rule diagnosis and result output module performs online or offline diagnosis according to the reviewed fault rules in the rule configuration library.

[0078] In the system implementation, the system mainly includes a device parsing and data acquisition module, a time processing and event construction module, a deep feature acquisition module, a feature label generation module, a coupling and evidence aggregation module, a rule diagnosis and result output module, a rule calibration module, and a rule configuration library.

[0079] The equipment analysis and data acquisition module receives the equipment identifier, the target diagnostic object identifier, and the diagnostic time, analyzes the equipment hierarchy and the affiliation of the measurement points, and obtains mechanical monitoring data and process monitoring data from sensors, condition monitoring systems, distributed control systems, data acquisition and monitoring systems, or historical databases.

[0080] The time processing and event construction module performs time correction, engineering unit unification, resampling, and data validity judgment on different data sources. It identifies abnormal vibration events of each shaft vibration channel and process events of each process parameter, and constructs multi-channel vibration events based on bearing position, detection direction, event start time difference, event overlap, and the number or proportion of abnormal channels.

[0081] The deep feature acquisition module selects a representative moment based on the type of shaft vibration anomaly event, retrieves the corresponding original waveform, and extracts the waveform spectrum features and shaft center trajectory features, or receives the waveform spectrum features and shaft center trajectory features provided by an external condition monitoring system.

[0082] The feature label generation module encapsulates features from different sources into standard feature samples and generates feature labels based on the judgment thresholds and feature combination conditions in the rule configuration library.

[0083] The coupling and evidence aggregation module calculates the initial time difference, event interval overlap ratio, peak time difference, and recovery order between the process event to be associated and the multi-channel vibration event, generates process-vibration coupling evidence, and performs source deduplication and hierarchical evidence aggregation based on the original event identifier.

[0084] The rule diagnosis and result output module sequentially performs key data checks, necessary condition judgments, directional timing condition judgments, exclusionary counter-evidence judgments, supporting item bonuses and ordinary counter-evidence item deductions to form fault scores and evidence levels, and outputs the main fault, candidate faults and corresponding handling suggestions.

[0085] The rule calibration module utilizes historical cases with fault truth value labels to perform offline calibration of initial fault rules established based on fault mechanisms and expert experience. Fault truth values ​​can be classified according to their reliability level. Among them, results confirmed by disassembly and inspection, component replacement, or shutdown testing have high reliability; consistent conclusions reached by field experts based on complete spectrum, shaft center trajectory, and process logs have the next highest reliability; results inferred solely from alarm records or single signals are used to assist in analysis.

[0086] During rule calibration, calibration data and independent verification data can be divided according to the generating unit to avoid adjacent time slices of the same generating unit from simultaneously entering calibration and verification data. The calibration content includes judgment thresholds, necessary conditions, supporting conditions, counter-evidence conditions, and scoring weights. The calibrated and approved fault rules are written into the rule configuration library. During actual online or offline diagnosis, the rules in the rule configuration library are executed, and the rules are not changed automatically during the diagnosis process.

[0087] During verification, the method of this invention can be compared with the single vibration threshold method, the multi-parameter weighted method without setting time coupling conditions, and the method using only the spectrum or shaft center trajectory. The identification results, false alarms and missed alarms of various faults are statistically analyzed. The verification results are used to adjust the judgment threshold and scoring parameters, and the predetermined accuracy rate is not preset in the absence of real test data.

[0088] The following provides supplementary explanations for five typical diagnostic scenarios: In the oil film whirl diagnostic scenario, the system can meet the necessary conditions for oil film whirl when it repeatedly detects low-frequency components of 0.38X to 0.48X at multiple adjacent representative moments of the same bearing, and simultaneously detects a decrease in the repeatability of the shaft center trajectory, a change in the precession direction or a change in the trajectory shape, and detects corresponding changes in lubricating oil temperature, lubricating oil pressure or bearing temperature. When the inlet flow rate, intake pressure and anti-surge valve opening do not form a directional coupling with the multi-channel vibration event, this fact can serve as counter-evidence to exclude surge, giving priority to oil film whirl in the candidate ranking.

[0089] As a parameter example, when the same bearing repeatedly exhibits a frequency band characteristic of 0.38X to 0.48X at adjacent representative moments, the shaft center trajectory repeatability score is less than 0.45, and the lubricating oil temperature is higher than 48 degrees Celsius, a supporting chain of evidence for oil film eddy can be formed. The specific threshold is determined based on the historical baseline of the corresponding unit.

[0090] In unbalanced fault diagnosis scenarios, when the system detects that the vibration remains high for a long time window, the 1X component dominates the waveform spectrum, the shaft center trajectory area increases without significant elongation, and the inlet flow rate, intake pressure, and anti-surge valve opening remain stable before and after the anomaly occurs, an evidence chain for unbalanced mechanical faults can be formed.

[0091] As an example of parameters, when vibration consistently exceeds 35 micrometers within a 30-day trend window, the 1X dominance ratio exceeds 0.7, the axis trajectory area exceeds 1300 square micrometers, and the trajectory axis-to-length ratio is less than 2.2, these characteristics are considered supporting evidence for imbalance-type faults. If the process event and the vibration event form a recurring directional coupling, the evidence level for imbalance-type faults is reduced, and airflow disturbance, rotational stall, or surge is further evaluated.

[0092] In blade damage diagnosis scenarios, when multiple shaft vibration channels experience synchronous jumps within a short period of time, the 1X phase undergoes a sudden change, and the vibration remains at a high level for a period of time after the jump, while process parameters such as inlet flow rate and inlet pressure do not show continuous periodic fluctuations, this can form candidate evidence of blade damage or foreign object intrusion.

[0093] As a parameter example, when a multi-channel shaft vibration exhibits a jump of more than 10 micrometers within the synchronization tolerance, a 1X phase change of more than 20 degrees, and the vibration remains at a high level after the jump, it is considered as supporting evidence of blade damage. If the anomaly occurs only in a single channel and is accompanied by clipping, constant value, over-range, or quality code anomalies, then the isolated anomaly of a single channel is considered as counter-evidence of blade damage, and sensor anomalies are evaluated first.

[0094] In the scenario of airflow disturbance and surge precursor diagnosis, when the inlet flow rate drops significantly relative to the pre-event baseline, and then multiple channels fluctuate rapidly and frequently within the directional allowable hysteresis range set by the corresponding fault rule, while the inlet pressure fluctuates at a low frequency periodic rate, and the anti-surge valve exhibits rapid or repetitive action beyond the control dead zone, surge or surge precursor candidate is formed.

[0095] As an example of parameters, a decrease in inlet flow rate of more than 20% relative to the pre-event baseline can be defined as a significant decrease in inlet flow rate, and a fluctuation score of 1.2 can be used as a criterion for a rapid fluctuation event. If the process event and the multi-channel vibration event meet the conditions of start time difference, event overlap ratio, or peak time difference, then supporting evidence for surge precursors is formed. If there is only directional coupling between the process event and the multi-channel vibration event, but no periodic fluctuations or repetitive events in the inlet pressure, then the candidate result is downgraded to airflow disturbance.

[0096] In the scenario of distinguishing between rotating stall and surge, when the inlet flow rate decreases and there are continuous or intermittent subsynchronous components or low-frequency aerodynamic components, accompanied by fluctuations in some rotor shaft vibration channels, but without strong periodic flow and pressure reversal swings throughout the machine or repeated large-scale movements of the anti-surge valve, rotating stall is considered the preferred candidate.

[0097] When multiple channels of vibration, periodic changes in inlet flow rate, reverse oscillation of intake pressure, and repeated operation of the anti-surge valve occur simultaneously, and each event meets the corresponding directional coupling conditions, surge is considered the priority candidate. When the low-frequency components are mainly concentrated in the 0.38X to 0.48X frequency band, and the evidence from the shaft center trajectory and lubrication parameters is concentrated in a single bearing, it is not directly identified as rotational stall or surge, but rather evaluated using the oil film whirl rule.

[0098] The specific time windows, synchronization tolerances, feature thresholds, scoring parameters, and evidence level thresholds in the above embodiments are merely examples provided for the purpose of illustrating the present invention. In practical applications, these settings and adjustments can be made based on the centrifugal compressor's structural form, rated speed, load range, inlet medium, sensor configuration, and historical operating data.

[0099] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A fault diagnosis method for a centrifugal compressor, characterized in that, Includes the following steps: S1: Obtain the equipment level information of the target centrifugal compressor, determine the measuring points, bearings and rotors corresponding to the target diagnostic object based on the equipment level information, and establish the membership relationship between the measuring points, bearings, rotors and the target centrifugal compressor; S2: Obtain the mechanical monitoring data and process monitoring data of the target centrifugal compressor, convert the timestamps of different data sources into a unified event time, and determine the validity of each monitoring data based on time quality and data quality; S3: Identify abnormal shaft vibration events from the shaft vibration trend sequence; construct multi-channel vibration events based on the bearing position, detection direction, and time relationship between abnormal shaft vibration events corresponding to different shaft vibration channels; identify process events from the process monitoring data respectively; S4: Select a representative moment for the identified shaft vibration abnormal event, and obtain the waveform spectrum features and shaft center trajectory features corresponding to the representative moment; S5: Convert trend features, multi-channel vibration events, process events, waveform spectrum features, and shaft trajectory features into feature labels with fault physical meaning, and associate feature labels generated from the same abnormal event with the same original event identifier; S6: Based on the preset centrifugal compressor fault rules, calculate the start time difference, event interval overlap ratio, peak time difference and recovery order between process events and multi-channel vibration events, and generate process-vibration coupling evidence based on the calculation results and the directional timing conditions configured according to the corresponding fault rules. S7: Deduplicatize the feature tags and process-vibration coupling evidence from the same source, and aggregate the evidence according to the hierarchical order of measuring point, bearing, rotor and target centrifugal compressor; S8: Filter valid candidate faults based on the key data, necessary conditions, directional timing conditions and exclusionary counter-evidence conditions required for each candidate fault. Score the valid candidate faults based on the supporting conditions and ordinary counter-evidence conditions. Determine the main fault and candidate faults based on the score results and the exclusion relationship between the candidate faults. Output the corresponding fault evidence and handling suggestions.

2. The fault diagnosis method for a centrifugal compressor according to claim 1, characterized in that, In S2, the mechanical monitoring data includes at least shaft vibration data from multiple shaft vibration channels, and includes at least one of shaft displacement data, bearing temperature data, lubricating oil temperature data, and rotational speed data; the process monitoring data includes at least one of inlet flow rate, inlet pressure, inlet temperature, and anti-surge valve opening; key data and auxiliary data are configured for different candidate faults, and when key data is missing or invalid, the corresponding candidate fault is excluded from the valid candidate set; when auxiliary data is missing or invalid, the supporting conditions corresponding to the auxiliary data are stopped or the evidence level of the corresponding candidate fault is reduced.

3. The fault diagnosis method for a centrifugal compressor according to claim 1, characterized in that, In S2, converting timestamps from different data sources into a unified event time includes the following steps: The system acquires the data source timestamp, acquisition timestamp, sampling period, engineering unit, and quality code of various monitoring data. It then corrects the data source timestamp according to the unified clock on the equipment side and records the clock deviation of the data source clock relative to the unified clock on the equipment side. The original sampling period of each monitoring data is retained, and the shaft vibration trend data and process trend data are resampled to a unified time grid. For continuous data, previous value hold or interpolation is used, and for discrete state data, previous value hold is used. When the continuous absence of measurement exceeds the absence tolerance of the corresponding data source, the quality code is invalid, the clock deviation exceeds the preset time correction tolerance, or the engineering unit cannot reach a consensus, the corresponding monitoring data will be marked as unavailable, and the use of unavailable monitoring data to meet the necessary conditions for candidate faults will be prohibited.

4. The fault diagnosis method for a centrifugal compressor according to claim 1, characterized in that, In S3, constructing multi-channel vibration events includes the following steps: Identify abnormal shaft vibration events from each shaft vibration channel, and determine the start time, end time, and duration of each abnormal shaft vibration event; When the maximum difference in the start time of each abnormal shaft vibration event within the same channel group does not exceed the synchronization tolerance, and the number of abnormal channels reaches a preset number or the proportion of abnormal channels to effective channels reaches a preset proportion, the corresponding abnormal shaft vibration events are merged into a multi-channel vibration event. The channel group includes at least one of the following: two orthogonal direction shaft vibration channel groups for the same bearing, a shaft vibration channel group for the same rotor, and a whole-machine shaft vibration channel group for the target centrifugal compressor. The shaft vibration channel group for the same rotor includes two orthogonal direction shaft vibration channels corresponding to the first bearing position and the second bearing position, respectively. The first bearing position and the second bearing position are the connecting end bearing position and the non-connecting end bearing position of the rotor, respectively. When both orthogonal direction shaft vibration channels are effective and both exhibit abnormal shaft vibration events within the synchronization tolerance, a four-channel synchronous event is generated.

5. The fault diagnosis method for a centrifugal compressor according to claim 1, characterized in that, In S4, a representative moment is selected for the identified shaft vibration anomaly event, including the following steps: The corresponding time selection rule is selected according to the event type of shaft vibration abnormal event. Specifically, for abrupt type shaft vibration abnormal events, the time when the sudden change occurs is selected; for continuous fluctuation type shaft vibration abnormal events, the time when the fluctuation is the greatest is selected; and for stable high-level type shaft vibration abnormal events, the time within the stable abnormal range is selected. The original waveform data of the corresponding shaft vibration channel is retrieved according to the representative moment, and the waveform spectrum characteristics are calculated based on the original waveform data; synchronous original waveform data of the shaft vibration channels in two orthogonal directions of the same bearing are retrieved, and the shaft center trajectory is generated and the shaft center trajectory characteristics are extracted based on the synchronous original waveform data of the two orthogonal directions; or the waveform spectrum characteristics and shaft center trajectory characteristics generated by the external monitoring system according to the representative moment are received; wherein, the waveform spectrum characteristics include at least one of order spectrum characteristics, low frequency characteristics, phase characteristics, clipping characteristics, glitch characteristics, peak-valley asymmetry characteristics and baseline drift characteristics, and the shaft center trajectory characteristics include at least one of precession direction, trajectory repeatability, trajectory axis length ratio, trajectory area and trajectory shape; wherein, the fundamental frequency of the order spectrum is determined based on the rotational speed data verified by data quality.

6. The fault diagnosis method for a centrifugal compressor according to claim 1, characterized in that, In S5, trend features, multi-channel vibration events, process events, waveform spectrum features, and shaft trajectory features are encapsulated into standard feature samples with measurement point identifiers, bearing identifiers, rotor identifiers, event times, original event identifiers, quality codes, and feature values, respectively. The standard feature samples are then converted into corresponding feature labels according to preset thresholds and feature combination conditions. For multiple feature labels generated from the same abnormal event, the same source deduplication is performed at the measurement point layer according to the original event identifier, and when aggregating step by step to the bearing layer, rotor layer and target centrifugal compressor layer, the same abnormal event is avoided from being counted repeatedly according to the original event identifier; For the same abnormal event, feature labels generated in different trend time windows are retained, and the feature labels are merged into a set of evidence in the same fault rule.

7. The fault diagnosis method for a centrifugal compressor according to claim 1, characterized in that, In S6, generating evidence of process-vibration coupling includes the following steps: Identify the corresponding process events from the trend data of inlet flow rate, inlet pressure, inlet temperature, and anti-surge valve opening, respectively; Based on the order of occurrence and duration of process events, multiple related process events are combined into a composite process event, and a single process event or the composite process event is identified as a process event to be associated. Calculate the start time difference, event interval overlap ratio, peak time difference, and recovery order of the process event to be associated relative to the multi-channel vibration event; When the initial time difference is within the directional allowable lag interval set by the corresponding fault rule, and the event interval overlap ratio, peak time difference, or recovery order meets the corresponding judgment conditions, the process-vibration coupling evidence is generated.

8. The fault diagnosis method for a centrifugal compressor according to claim 1, characterized in that, In S6, the centrifugal compressor fault rules include key data conditions, necessary conditions, supporting conditions, ordinary counter-evidence conditions, exclusionary counter-evidence conditions, and fault exclusion relationships. The necessary conditions are formed by combining feature tags and process-vibration coupling evidence according to AND logic, OR logic, logic that satisfies at least a preset number, or logic that satisfies at least a preset proportion. When the key data is valid, the necessary conditions are met, the directional timing conditions are met, and the exclusionary counter-evidence conditions are not hit, optional bonus points are determined according to the supporting conditions, and counter-evidence deduction points are determined according to the ordinary counter-evidence conditions. The candidate fault scores are obtained under the limit of the optional bonus points and the maximum total score of the fault rules. The scores of different candidate faults are normalized. The main fault and candidate faults are determined according to the normalized scores, fault exclusion relationships, and preset fault priorities. The corresponding fault evidence is retained for faults that can coexist.

9. A fault diagnosis system for a centrifugal compressor, for implementing the method according to any one of claims 1 to 8, characterized in that, include The equipment analysis and data acquisition module is used to acquire equipment hierarchy information, mechanical monitoring data and process monitoring data of the target centrifugal compressor, and to establish the hierarchical relationship between measuring points, bearings, rotor and the target centrifugal compressor. The time processing and event construction module is used to convert timestamps from different data sources into unified event times, determine the validity of various monitoring data based on time quality and data quality, identify shaft vibration abnormal events from shaft vibration trend sequences and construct multi-channel vibration events, and identify process events from process monitoring data. The deep feature acquisition module is used to select a representative moment for the shaft vibration abnormal event and acquire the waveform spectrum features and shaft center trajectory features corresponding to the representative moment. The feature label generation module is used to convert trend features, multi-channel vibration events, process events, waveform spectrum features, and shaft center trajectory features into feature labels, and associate feature labels generated from the same abnormal event with the same original event identifier. The coupling and evidence aggregation module is used to generate process-vibration coupling evidence based on the start time difference, event interval overlap ratio, peak time difference and recovery order between process events and multi-channel vibration events. The module performs deduplication of the feature tags and process-vibration coupling evidence, and aggregates the evidence according to the hierarchical order of measuring point, bearing, rotor and target centrifugal compressor. The rule diagnosis and result output module is used to screen valid candidate faults based on key data, necessary conditions, directional timing conditions and exclusionary counter-evidence conditions, score valid candidate faults based on supporting conditions and ordinary counter-evidence conditions, determine the primary fault and candidate faults based on the scoring results and fault exclusion relationships, and output corresponding fault evidence and handling suggestions.

10. The fault diagnosis system for a centrifugal compressor according to claim 9, characterized in that, It also includes a rule calibration module and a rule configuration library; the rule calibration module is used to perform offline calibration of the judgment threshold, necessary conditions, supporting conditions, counter-evidence conditions and scoring weights in the initial fault rules established by fault mechanism and expert experience based on historical fault samples with fault truth value labels, and write the approved fault rules into the rule configuration library; the rule configuration library is used to configure corresponding fault rules according to the centrifugal compressor model, historical baseline of measuring points and current operating conditions, and the rule diagnosis and result output module performs online diagnosis or offline diagnosis according to the approved fault rules in the rule configuration library.

Images