A valve real-time flow monitoring method based on big data analysis

Abstract

The application discloses a valve real-time flow monitoring method based on big data analysis, comprising: collecting valve multi-source working condition data and preprocessing to form working condition data set for modeling; constructing a digital twin model containing physical and correction units, outputting predicted flow and residual error; sliding statistics on residual error sequence, self-calibration update under steady-state working condition when exceeding limit; extracting residual time-frequency domain features, combining historical abnormal templates to generate residual amplification factor and score; comparing abnormal score with working condition cluster quantile threshold, outputting early warning or alarm and labeling abnormal type; performing order-preserving deformation and working condition cluster mapping, combining conformal calibration and playback verification to complete online adaptation. The application realizes high-precision prediction of valve flow, weak abnormality identification and cross-working condition real-time monitoring by introducing digital twin self-calibration, residual amplification factor enhancement and working condition cluster isomorphic adaptation mechanism.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of industrial process detection and control, and particularly relates to a valve real-time flow monitoring method based on big data analysis. BACKGROUND

[0002] As a key regulating element widely used in process industry, the flow monitoring of valve is of great significance to ensure process safety, optimize energy utilization and realize intelligent operation. The existing valve flow monitoring methods mainly rely on direct measurement by flow meter or indirect calculation of flow rate by using the relationship between valve opening and differential pressure. In the direct measurement method, common flow meters include vortex flow meter, orifice flow meter and electromagnetic flow meter, etc. This kind of method can accurately obtain instantaneous flow data under ideal working conditions, but in the long-term operation process, the flow meter is often affected by sediment, wear, corrosion and medium disturbance, and problems such as zero drift, low measurement accuracy and high maintenance cost are prone to occur. The calculation method based on valve opening and differential pressure usually relies on the opening-flow characteristic curve calibrated at the factory, and the flow rate is derived by using a fixed empirical formula. However, this method has poor adaptability under different medium conditions and operating conditions, and once the valve performance is aged or the operating condition is changed, the prediction result often deviates greatly.

[0003] With the development of industrial big data and digital technology, some researches try to introduce data-driven models or digital twin technology for valve flow prediction, and establish data regression models or simulation models to improve the accuracy. However, the existing methods generally have three shortcomings. First, in the long-term operation process, the parameters of the digital twin model are often fixed and cannot be automatically updated according to the residual error, which leads to the gradual degradation of prediction accuracy over time. Second, for weak anomalies such as leakage, blockage and hysteresis, the existing threshold-based discrimination method has insufficient sensitivity, and often needs a significant deviation to trigger an early warning, which is easy to miss the key anomalies. Finally, the model transferability across valves and operating conditions is limited, and the existing methods are difficult to quickly adapt to different types of valves or different operating conditions, which limits the application.

[0004] The existing valve flow monitoring technology cannot meet the needs of high precision, strong robustness and cross-condition scalability in industrial field.

[0005] Therefore, how to provide a valve real-time flow monitoring method based on big data analysis is a problem to be solved by those skilled in the art. SUMMARY

[0006] One purpose of the present application is to propose a valve real-time flow monitoring method based on big data analysis, which fully utilizes digital twin modeling, self-calibration updating mechanism, residual amplification factor enhancement mechanism, and cross-valve working condition order-preserving parameter deformation and isomorphic mapping adaptation mechanism, and details the prediction, anomaly detection and adaptive correction process of valve flow under different operating conditions, with the advantages of long-term operation precision, sensitive identification of weak anomalies, and strong cross-valve and cross-condition promotion capability.

[0007] According to the valve real-time flow monitoring method based on big data analysis, the method comprises the following steps:

[0008] Collect valve flow data, preprocess the valve flow data to form a working condition data set;

[0009] Based on the valve characteristic parameters and medium physical property parameters in the working condition data set, a digital twin model containing a physical calculation unit and a correction calculation unit is constructed, the valve predicted flow is output by using the digital twin model, and the residual between the actual flow and the predicted flow is calculated to obtain a residual sequence;

[0010] The residual sequence is statistically analyzed by using a sliding time window, and it is judged whether the residual mean and variance exceed the preset threshold value, and when the over-limit condition is met, the update of the digital twin model parameters is executed under the steady state working condition, and the updated valve predicted flow and residual sequence are output;

[0011] The updated residual sequence is extracted in time domain and frequency domain, compared with the three types of feature vectors of leakage, blockage and hysteresis in the historical abnormal template library, a residual amplification factor is generated, and the abnormal score is obtained after the residual sequence is enhanced by the residual amplification factor;

[0012] The abnormal score is compared with the quantile threshold value of the corresponding working condition cluster, and the pre-warning information or alarm information is output according to the comparison result, and the abnormal type label is marked;

[0013] Based on the working condition data set, the valve characteristic parameters, the residual features and the abnormal score, the order-preserving parameter deformation and the working condition cluster isomorphic mapping adaptation mechanism are executed on the big data platform, the monotone spline deformation is performed on the factory opening-flow characteristic curve and the residual feature distribution of the source valve, the isomorphic mapping of the quantile is kept for the working condition cluster, the conformal calibration is performed on the valve predicted flow and the residual sequence, and the deformation parameters and the threshold value are issued after verification by the playback set on the edge node, and the new valve or new working condition is adapted online after verification, and the original parameters are kept and recorded when the verification fails.

[0014] Optionally, the valve flow data includes valve opening data, valve inlet and outlet pressure data, valve differential pressure data, medium temperature data, actuator current data, vibration signal data, acoustic signal data, and flow data measured by a flowmeter.

[0015] Optionally, the preprocessing of the valve flow data includes time alignment processing, noise filtering processing, outlier rejection processing, and missing data completion processing on the collected valve flow data.

[0016] Optionally, the working condition cluster refers to, under the conditions of fixed valve model and medium category, partitioning and clustering valve flow data according to four key characteristics of valve opening, valve differential pressure, medium temperature, and medium property to form a steady-state data set with boundary constraints, and each working condition cluster is represented by a characteristic boundary and a unique identifier.

[0017] Optionally, the residual sequence is obtained by:

[0018] Reading data in the working condition data set, constructing a digital twin model and instantiating it in an edge computing node, the digital twin model being composed of a physical computing unit and a correction computing unit, taking data in the working condition data set as input and running at a uniform sampling period;

[0019] Constructing a physical computing unit, establishing a monotonic spline mapping based on the factory opening-flow characteristic curve, parameterizing the factory opening-flow characteristic curve with a plurality of opening-flow characteristic control points, introducing a medium property correction to the differential pressure-flow relationship, and imposing monotonicity and smoothness constraints, and outputting a time-aligned physical predicted flow sequence;

[0020] Constructing a correction computing unit, selecting time series features of working condition data within a preset historical window length as input data, adopting a fixed-layer feedforward neural network as a data-driven structure, normalizing the input data, imposing an amplitude limiting constraint on the output result, generating a time-aligned correction amount sequence and corresponding confidence indication information;

[0021] Adding and fusing the physical predicted flow sequence and the correction amount sequence to generate a valve predicted flow sequence, performing physical feasibility verification and boundary truncation, and generating prediction uncertainty indication information according to physical constraints and confidence indication information;

[0022] Determining the actual flow sequence, when a flowmeter is available on site, taking the verified flowmeter measurement value as the actual flow sequence, when a flowmeter is not available on site, performing differential pressure method flow calculation based on the valve, differential pressure, and medium property to form the actual flow sequence, aligning the actual flow sequence and the valve predicted flow sequence, and then subtracting them hour by hour to obtain the residual sequence and outputting it.

[0023] Optionally, the output of the updated valve predicted flow and residual sequence includes:

[0024] Set the length and step of the sliding time window, and divide the residual sequence into windows in chronological order, and calculate the residual mean and residual variance in each sliding time window;

[0025] Set the mean threshold and variance threshold for each sliding time window, and determine the steady state condition in the same sliding time window, which includes the valve opening rate, differential pressure change rate and medium temperature fluctuation of multi-source working condition data not exceeding the pre-defined value;

[0026] When the residual mean or residual variance of any sliding time window exceeds the corresponding threshold and the steady state condition is met, start the double evidence trigger of digital twin model parameter update, which includes residual statistical over-limit evidence and multi-source working condition data consistency evidence;

[0027] After the double evidence trigger, perform local order-preserving deformation update on the physical calculation unit, taking the discrete control points of the opening-flow characteristic curve as the update object, and only deform in the opening interval corresponding to the trigger sliding time window, maintain the monotonicity and smoothness of the curve, and limit the single deformation amplitude not to exceed the preset proportion;

[0028] Perform restricted fine tuning on the correction calculation unit, adjust the output side parameters, use a segmented weighting strategy based on residual amplitude to suppress the influence of abnormal samples, set the maximum step and output range constraints, and generate the updated valve predicted flow sequence;

[0029] Construct a shadow twin for comparison and verification, select the historical data segment adjacent to the trigger sliding time window for rapid verification, and the verification content includes two indicators of residual mean and residual variance, and if the verification is passed, write the update to the effective version and output the updated valve predicted flow and corresponding residual sequence.

[0030] Optionally, the generation of the residual amplification factor includes:

[0031] Receive the updated residual sequence, set the length and step of the sliding window, and divide the residual sequence into consecutive analysis windows in chronological order;

[0032] Calculate the time domain statistics in each analysis window, including the mean, standard deviation, skewness and kurtosis of the residual sequence, to form a time domain feature vector;

[0033] Perform spectral analysis on the residual sequence in each analysis window to extract low, medium and high frequency energy, main peak frequency and spectral entropy, and form a frequency domain feature vector;

[0034] The time domain feature vector and the frequency domain feature vector are subjected to similarity calculation with feature vectors corresponding to the leakage template, the blockage template and the delay template in the historical abnormal template library, to obtain similarity indexes of each abnormal type, and the maximum value is selected as the maximum similarity of the current analysis window;

[0035] The standard deviation, the selected band energy and the spectral entropy are normalized according to the corresponding historical baseline of the working condition cluster, and are combined with the maximum similarity by preset weight, and the combination result is limited in a preset range, to generate a residual amplification factor;

[0036] The baseline normalized average of the residual amplitude in the analysis window is multiplied by the residual amplification factor to obtain an abnormal score sequence.

[0037] Optionally, the abnormal score is compared with the quantile threshold value of the corresponding working condition cluster, and according to the comparison result, a pre-warning information or an alarm information is output, and an abnormal type label is marked, including:

[0038] According to the working condition characteristics, the working condition cluster to which the working condition belongs is determined, and the historical abnormal score distribution and the historical residual amplification factor distribution of the working condition cluster are called as the reference data for the current window determination;

[0039] Two preset quantile threshold values are calculated on the historical abnormal score distribution, which are respectively used as a pre-warning threshold value and an alarm threshold value;

[0040] The abnormal score of the current window is compared with the pre-warning threshold value and the alarm threshold value, to obtain one of three determination levels of normal, pre-warning or alarm;

[0041] The pre-warning duration window number and the alarm duration window number and the back-off allowance are set, and the determination levels of the continuous windows are accumulated and counted, when the accumulation reaches the corresponding duration window number, a pre-warning event or an alarm event is generated, and when the abnormal score of the continuous window is lower than the pre-warning threshold value minus the back-off allowance, the generated pre-warning or alarm state is cleared;

[0042] The three types of similarity indexes of leakage, blockage and delay are read, the abnormal type with the highest similarity value is selected as the abnormal type label of the current window, and is associated with the abnormal score, the residual amplification factor and the working condition cluster identification of the current window;

[0043] An event record is generated, the event record is written into an edge side event queue and is synchronized to a cloud end log, and the event count and the latest occurrence time of the corresponding working condition cluster are updated.

[0044] Optionally, based on the working condition data set, the valve characteristic parameter, the residual feature and the abnormal score, an order-preserving parameter morphing and working condition cluster isomorphic mapping adaptation mechanism is executed on a big data platform, including:

[0045] Determine the source valve and the target valve, read the working condition data set, the opening-flow characteristic control point, the working condition cluster set, the residual characteristic baseline, and the early warning threshold and the alarm threshold;

[0046] Perform the order-preserving parameter deformation on the opening-flow characteristic control point, the deformation is limited within the target valve working opening interval and remains monotonic and smooth, limit the single deformation amplitude, and generate the initial control point of the target valve;

[0047] According to four types of characteristics of the valve opening, the valve differential pressure, the medium temperature and the medium physical property, establish one-to-one mapping of the source side working condition cluster and the target side working condition cluster, and obtain the working condition cluster set of the target valve;

[0048] According to the quantile retention rule, convert the source side residual characteristic baseline and the quantile threshold into the residual characteristic baseline and the quantile threshold of the target valve, and update the frequency band energy, the standard deviation and the spectral entropy baseline related to the residual amplification factor synchronously;

[0049] Collect the working condition data of the target valve within the adaptation period, perform the predicted flow and residual calculation by using the initial control point and the quantile threshold of the target valve, and perform the conformal calibration by combining the residual characteristics and the abnormal score, and obtain the calibration control point and the calibration threshold of the target valve;

[0050] In the edge computing node, the calibration control point and the calibration threshold of the target valve are applied to the playback data, and the improvement results of the window residual mean and the window residual variance are compared, when the preset judgment condition is met, the adaptation online is completed, and when the preset judgment condition is not met, the time and the reason are recorded.

[0051] The beneficial effects of the present application are:

[0052] The present application introduces a self-calibration updating mechanism in the valve digital twin model, which can dynamically correct the parameters according to the residual changes in long-term operation, thereby effectively avoiding the precision decay problem of the traditional method under the condition of fixed parameters. By comparing the residual sequence of the real-time predicted flow and the actual flow, and combining the statistical judgment of the sliding time window, the present application can realize the automatic updating of the model parameters under the steady state working condition, so that the valve flow prediction result can keep high precision and stability for a long time.

[0053] The present application constructs a residual amplification factor enhancement mechanism, extracts statistical features of the updated residual sequence in the time domain and the frequency domain, and compares the similarity with the historical abnormal template to generate the amplified abnormal score, which realizes the sensitive identification of weak abnormal signals. Compared with the existing method which depends on fixed threshold, the present application can significantly improve the detection ability of early abnormalities such as leakage, blockage and hysteresis, which is helpful for the on-site operators to find potential risks in time and reduce the probability of operation accidents.

[0054] The application proposes an adaptation method across valves and across working conditions, which maps the characteristics of the source valve and the baseline parameters to the target valve through the order-preserving parameter deformation and working condition cluster isomorphism mapping mechanism, verifies and goes online in the adaptation period combined with conformal calibration, ensures that the monitoring method can still maintain reliability and accuracy under different types of valves and diversified working conditions, improves the popularization and application ability of the method, reduces the dependence on independent modeling and repeated calibration of each valve, and has strong practicability and expansibility. BRIEF DESCRIPTION OF DRAWINGS

[0055] The accompanying drawings are included to provide a further understanding of the application and are incorporated in and constitute a part of this specification, illustrate embodiments of the application and together with the description serve to explain the application, and do not limit the application. In the drawings:

[0056] Figure 1 A flowchart of a valve real-time flow monitoring method based on big data analysis proposed by the application;

[0057] Figure 2 A processing flowchart of a residual amplification factor enhancement mechanism of a valve real-time flow monitoring method based on big data analysis proposed by the application. DETAILED DESCRIPTION

[0058] The application will now be described in further detail with reference to the drawings. These drawings are simplified schematic diagrams, and only illustrate the basic structure of the application in a schematic manner, and therefore only show the components related to the application.

[0059] Reference Figure 1 and Figure 2 A valve real-time flow monitoring method based on big data analysis, comprising:

[0060] Collecting valve flow data, preprocessing the valve flow data to form a working condition data set;

[0061] Based on the valve characteristic parameters and medium physical property parameters in the working condition data set, a digital twin model containing a physical calculation unit and a correction calculation unit is constructed, the valve predicted flow is output by using the digital twin model, and the residual between the actual flow and the predicted flow is calculated to obtain a residual sequence;

[0062] The residual sequence is subjected to sliding time window statistics, and it is judged whether the residual mean and variance exceed the preset threshold, when the over-limit condition is met, the update of the digital twin model parameters is executed under the steady state working condition, and the updated valve predicted flow and residual sequence are output;

[0063] The updated residual sequence is subjected to time domain and frequency domain feature extraction, and compared with three types of feature vectors of leakage, blockage and hysteresis in the historical abnormal template library to generate a residual amplification factor, and the residual sequence is enhanced by the residual amplification factor to obtain an abnormal score;

[0064] The abnormal score is compared with the quantile threshold of the corresponding working condition cluster, and warning information or alarm information is output according to the comparison result, and an abnormal type label is marked;

[0065] Based on the working condition data set, valve characteristic parameters, residual features and abnormal scores, a monotonic spline transformation and working condition cluster isomorphic mapping adaptation mechanism is performed on a big data platform, the monotonic spline transformation is performed on the factory opening-flow characteristic curve and residual feature distribution of the source valve, the working condition cluster is subjected to quantile-preserving isomorphic mapping, the valve predicted flow and residual sequence are subjected to conformal calibration, and the transformation parameters and threshold are issued after verification by the playback set on the edge node, and the new valve or new working condition is adapted online after verification, and the original parameters are kept and recorded when the verification fails.

[0066] In the embodiment, the valve flow data includes valve opening data, valve inlet and outlet pressure data, valve differential pressure data, medium temperature data, actuator current data, vibration signal data, acoustic signal data, and flow data measured by a flowmeter.

[0067] In the embodiment, the preprocessing of the valve flow data includes time alignment processing, noise filtering processing, abnormal point elimination processing and missing data completion processing on the collected valve flow data.

[0068] In the embodiment, the working condition cluster refers to clustering the valve flow data according to four types of key features of valve opening, valve differential pressure, medium temperature and medium physical properties under the condition of fixed valve model and medium category, forming a steady-state data set with boundary constraints, and each working condition cluster is represented by a feature boundary and a unique identifier.

[0069] In the embodiment, the residual sequence is obtained, including:

[0070] The data in the working condition data set is read, a digital twin model is constructed and instantiated on an edge computing node, the digital twin model is composed of a physical computing unit and a correction computing unit, and the data in the working condition data set is input and run at a uniform sampling period;

[0071] The physical calculation unit is constructed, a monotone spline mapping is established based on the factory opening-flow characteristic curve, the factory opening-flow characteristic curve is parameterized by a plurality of opening-flow characteristic control points, a medium property correction to the differential pressure-flow relationship is introduced, and monotonicity and smoothness constraints are applied, and a physical predicted flow sequence aligned with time is output, wherein the monotonicity and smoothness constraints include:

[0072] The function relationship between opening and flow is strictly monotonically increasing in the entire opening interval, and multi-value mapping is avoided;

[0073] The continuity of the first-order derivative and the second-order derivative between adjacent control points is limited, and the curve is smooth and has no abrupt change;

[0074] The derivative is constrained to be non-zero in the low opening interval, and the derivative is gradually reduced in the high opening interval, which prevents the occurrence of non-physical sharp growth, the low opening interval refers to the valve opening being not greater than 30% of the rated opening or the corresponding flow being not greater than 30% of the rated flow, and the high opening interval refers to the valve opening being not less than 70% of the rated opening or the corresponding flow being not less than 70% of the rated flow;

[0075] The correction calculation unit is constructed, the time sequence features of the working condition data in a preset historical window length are selected as input data, a fixed number of layers of a feedforward neural network is used as a data-driven structure, the input data is normalized, and the output result is subjected to an amplitude limiting constraint, a correction amount sequence aligned with time and corresponding confidence indication information are generated, wherein the amplitude limiting constraint includes:

[0076] The absolute value of the correction amount is constrained to be not more than a preset upper limit of the difference between the flowmeter measured value and the physical predicted flow;

[0077] The change rate of the correction amount is constrained to be not more than the change amplitude of the residual mean value in adjacent time windows;

[0078] The minimum amplitude limiting is applied to the correction amount in the low opening working condition, and the maximum amplitude limiting is applied to the correction amount in the high opening working condition, the low working condition refers to the valve opening being not greater than 30% of the rated opening or the corresponding flow being not greater than 30% of the rated flow, and the high working condition refers to the valve opening being not less than 70% of the rated opening or the corresponding flow being not less than 70% of the rated flow;

[0079] The physical prediction flow sequence and the correction quantity sequence are added and fused to generate a valve prediction flow sequence, physical feasibility verification and boundary truncation are performed, and prediction uncertainty indication information is generated according to physical constraints and confidence indication information, wherein the physical constraints refer to limiting the value range of the predicted flow to be not less than zero and not more than the rated maximum flow of the valve, limiting the change rate of the predicted flow to be not more than the maximum allowable adjustment rate designed by the valve, ensuring that the predicted flow is not less than the minimum opening and closing flow in the low opening interval, and ensuring that the predicted flow gradually tends to saturation with the opening in the high opening interval;

[0080] An actual flow sequence is determined, when a flowmeter is present on site, the verified flowmeter measurement value is taken as the actual flow sequence, when a flowmeter is not present on site, the actual flow sequence is formed by differential pressure method flow calculation according to the valve, differential pressure and medium properties, and the actual flow sequence and the valve prediction flow sequence are aligned, then the residual sequence is obtained by subtraction hour by hour and output.

[0081] The present application introduces a double-structure design of a physical calculation unit and a correction calculation unit in a digital twin model, which not only ensures that the valve flow prediction result is consistent with the physical law, but also enhances the adaptability to non-linear factors under complex working conditions through neural network correction. The physical unit adopts monotone spline mapping and derivative constraint, so that the opening-flow relationship remains stable and reasonable for a long time; the correction unit avoids excessive correction through amplitude limiting constraint, and applies differential control in high and low opening intervals, thereby improving the reliability of prediction. The physical feasibility verification and uncertainty output are introduced to improve the credibility of the model. Finally, by comparing with the measured or differential pressure method calculation result, a residual sequence is formed to provide high-quality input for self-calibration and abnormality identification. Compared with the prior art, the present application can maintain high-precision prediction in long-term operation, and has the ability of sensitive abnormality identification and cross-condition adaptation, thereby improving the engineering practicability of valve flow monitoring.

[0082] In the embodiment, the output updated valve prediction flow and residual sequence comprises:

[0083] The length and step of the sliding time window are set, the residual sequence is divided into windows in time sequence, and the residual mean and residual variance in each sliding time window are calculated respectively;

[0084] Mean threshold and variance threshold are set for each sliding time window, and a steady state condition is determined in the same sliding time window, the steady state condition includes that the valve opening rate of change, the differential pressure rate of change and the medium temperature fluctuation of multi-source working condition data do not exceed the pre-defined value;

[0085] When the residual mean or residual variance of any sliding time window exceeds the corresponding threshold and the steady state condition is established, the double-evidence trigger of digital twin model parameter update is started, and the double-evidence trigger includes residual statistical over-limit evidence and multi-source working condition data consistency evidence, wherein:

[0086] The residual statistical over-limit evidence refers to that the residual variance calculated in a preset window exceeds the corresponding quantile threshold, and the over-limit state continues to exist in a plurality of continuous windows;

[0087] The multi-source working condition data consistency evidence refers to that the multi-source working condition data collected in the same time window satisfy the physical constraint relationship and the energy conservation condition, and no sensor drift or data conflict occurs;

[0088] After the double-evidence triggering, a local order-preserving deformation update is performed on the physical calculation unit, the discrete control points of the opening-flow characteristic curve are taken as the update objects, the deformation is only performed in the opening interval corresponding to the triggered sliding time window, the curve monotonicity and smoothness are maintained, and the single deformation amplitude is limited to not exceed a preset proportion;

[0089] A limited fine-tuning is performed on the correction calculation unit, the output side parameters are adjusted, a segmented weighted strategy based on the residual amplitude is adopted to suppress the influence of abnormal samples, a maximum step size and an output range constraint are set, and an updated valve predicted flow sequence is generated, and the segmented weighted strategy based on the residual amplitude is specifically:

[0090] When the residual amplitude is lower than a first preset threshold, it is considered that the sample has high credibility, and a higher weight is given to participate in the parameter update;

[0091] When the residual amplitude is between the first preset threshold and a second preset threshold, it is considered that the sample has a certain deviation, and the weight is segmented and decreased according to the residual amplitude;

[0092] When the residual amplitude exceeds the second preset threshold, it is considered that the sample may be an abnormal point or a sensor error, and the weight is weakened to close to zero;

[0093] A shadow twin is constructed for comparison and verification, a historical data segment adjacent to the triggered sliding time window is selected for rapid verification, the verification content includes two indexes of residual mean value reduction and residual variance reduction, and when the verification is passed, the update is written into an effective version and the updated valve predicted flow and the corresponding residual sequence are output, and the construction of the shadow twin for comparison and verification is.

[0094] An edge computing node is constructed to construct a shadow twin identical to the digital twin model structure, load the updated parameters and independently run, without affecting the digital twin model;

[0095] The historical working condition data segment adjacent to the triggered sliding time window is selected and input into the shadow twin, the residual sequence of the predicted flow and the actual flow is calculated and compared with the original digital twin model result;

[0096] If the residual mean and residual variance of the shadow twin are lower than those of the digital twin model, it is determined that the verification is passed, the updated result is written into the effective version, and the updated valve predicted flow and corresponding residual sequence are output, otherwise the update is discarded and the original parameters are maintained.

[0097] The application introduces a sliding time window mechanism in residual sequence monitoring, and combines double-evidence triggering strategies of residual statistical overrun and multi-source working condition consistency, effectively avoiding misjudgment caused by single index fluctuation, thereby ensuring the reliability of parameter updating. In the model updating link, the physical calculation unit adopts local order-preserving deformation, and only the control points in the trigger interval are corrected by a limited amplitude, which ensures the monotony and smoothness of the opening-flow curve, and avoids the uncertainty brought by large-scale adjustment. The correction calculation unit introduces a segmented weighted fine-tuning strategy based on residual amplitude, strengthens the learning of credible samples, and weakens the weight of suspected abnormal samples, thereby improving the robustness. Before going online, a shadow twin is constructed for comparison and verification, and the effectiveness is verified by double-index verification of residual mean and variance, ensuring that only when the performance is improved can the main model be written. The application can maintain high-precision prediction in long-term operation, has self-adaptive, self-verification and abnormal suppression capabilities, and improves the reliability and practicality of valve flow monitoring.

[0098] In the embodiment, the generated residual amplification factor is used to enhance the residual sequence to obtain an abnormal score, which includes:

[0099] The updated residual sequence is received, the length and step of the sliding window are set, and the residual sequence is divided into continuous analysis windows in chronological order;

[0100] Time domain statistics are calculated in each analysis window, including the mean, standard deviation, skewness and kurtosis of the residual sequence, to form a time domain feature vector;

[0101] The residual sequence is subjected to spectral analysis in each analysis window, and low-frequency energy, medium-frequency energy and high-frequency energy, main peak frequency and spectral entropy are extracted to form a frequency domain feature vector;

[0102] The time domain feature vector and the frequency domain feature vector are subjected to similarity calculation with the feature vectors corresponding to the leakage template, the blockage template and the hysteresis template in the historical abnormal template library, to obtain similarity indexes of each abnormal type, and the maximum value is selected as the maximum similarity of the current analysis window, wherein:

[0103] The historical abnormal template library refers to the statistical features of the residual sequence in the time domain and the frequency domain, which are extracted from the abnormal working condition samples confirmed during the long-term operation of the valve, to form a standardized feature vector set;

[0104] The feature vector set is stored in a classified manner according to abnormal types, and contains a leakage abnormal template, a blockage abnormal template and a delay abnormal template;

[0105] The standard deviation, the selected band energy and the spectral entropy are normalized according to the corresponding working condition cluster historical baseline, and are combined with the maximum similarity by preset weight, the combination result is limited in a preset range, and a residual amplification factor is generated;

[0106] The baseline normalized average of the residual amplitude in the analysis window is multiplied by the residual amplification factor to obtain an abnormal score sequence.

[0107] The application realizes accurate identification of different types of abnormal patterns by introducing time domain and frequency domain double feature extraction in residual analysis and combining the similarity calculation of the historical abnormal template library. The time domain statistics can capture the overall offset and distribution characteristics of the residual, and the frequency domain features reflect the energy distribution of the vibration and periodic abnormality, which guarantees comprehensive perception of abnormal information. In the abnormal identification link, the residual amplification factor generation mechanism is innovatively proposed, the normalized results of the standard deviation, the band energy and the spectral entropy are combined with the maximum similarity by weight, and the result is limited in a range, thereby effectively avoiding misjudgment caused by noise or isolated abnormality. The abnormal score generated by weighting the normalized residual amplitude by the amplification factor not only improves the amplification detection capability of weak abnormalities, but also maintains the stability and robustness of the result. Compared with the traditional detection method based on fixed threshold, the application can significantly improve the detection rate of early abnormalities such as leakage, blockage and delay, and provides higher sensitivity and accuracy for real-time flow monitoring of valves.

[0108] In the embodiment, the comparison of the abnormal score with the quantile threshold of the corresponding working condition cluster, outputting the early warning information or the alarm information according to the comparison result, and labeling the abnormal type label, comprises:

[0109] According to the working condition characteristics, the belonging working condition cluster is determined, the historical abnormal score distribution and the historical residual amplification factor distribution of the working condition cluster are called as the reference data for the current window determination;

[0110] Two preset quantile thresholds are calculated on the historical abnormal score distribution, which are respectively used as the early warning threshold and the alarm threshold, and the calculation of the two preset quantile thresholds on the historical abnormal score distribution specifically comprises:

[0111] From the abnormal score sequence recorded in the long-term operation process of the valve, a complete historical distribution covering multiple working conditions is extracted;

[0112] A first quantile point is determined on the historical abnormal score distribution, which is used as the early warning threshold for identifying early potential abnormalities;

[0113] determine a second quantile point on the historical abnormal score distribution as an alarm threshold for identifying an abnormal working condition that has reached a serious degree;

[0114] compare the abnormal score of the current window with the pre-warning threshold and the alarm threshold to obtain one of three determination levels of normal, pre-warning or alarm;

[0115] set a pre-warning duration window number, an alarm duration window number and a back-off margin, accumulate statistics of determination levels of consecutive windows, generate a pre-warning event or an alarm event when the accumulated statistics reaches the corresponding duration window number, and clear the generated pre-warning or alarm state when the abnormal score of the consecutive window is lower than the pre-warning threshold minus the back-off margin, wherein the back-off margin refers to a safety buffer interval set when the abnormal score is lower than the pre-warning threshold by a certain margin, and is used to avoid frequent switching of pre-warning or alarm states caused by critical fluctuations of the score;

[0116] read the three similarity indexes of leakage, blockage and delay, select the abnormal type with the highest similarity value as the abnormal type label of the current window, and associate the abnormal type label with the abnormal score, the residual amplification factor and the working condition cluster identifier of the current window;

[0117] generate an event record containing a timestamp, a determination level, an abnormal type label, an abnormal score, a residual amplification factor and a working condition cluster identifier, write the event record into an edge-side event queue and synchronize the event record to a cloud log, and update the event count and the latest occurrence time of the corresponding working condition cluster.

[0118] The application introduces a working condition cluster division mechanism, so that abnormal determination can be based on comparison of historical score distribution and residual amplification factor distribution under similar working conditions, effectively avoiding the failure problem of a single fixed threshold under cross-working condition. In threshold setting, a quantile point adaptive calculation method is used to dynamically generate a pre-warning threshold and an alarm threshold, ensuring the flexibility and accuracy of valve abnormality detection under different working conditions. In the event triggering link, a pre-warning duration window number, an alarm duration window number and a back-off margin are set, and through cumulative statistics and a safety buffer mechanism, frequent false alarms caused by critical fluctuations of the score are effectively suppressed. The application combines three similarity indexes of leakage, blockage and delay, automatically labels the abnormal type, and writes the abnormal score, the amplification factor and the working condition cluster identifier into the event record, realizing edge-side rapid determination and cloud log synchronization. Innovatively, the distribution threshold, adaptive determination and abnormal type labeling are combined together, not only improving the accuracy and stability of valve flow monitoring, but also enhancing the practicability of the system under long-term operation and complex working conditions.

[0119] In the embodiment, based on the working condition data set, the valve characteristic parameter, the residual feature and the abnormal score, a sequence-preserving parameter morphing and working condition cluster isomorphic mapping adaptation mechanism is performed on a big data platform, including:

[0120] determining the source valve and the target valve, reading the working condition data set, the opening-flow characteristic control point, the working condition cluster set, the residual characteristic baseline and the pre-warning threshold and the alarm threshold;

[0121] performing a monotonic and smooth parameter transformation on the opening-flow characteristic control point, limiting the single transformation amplitude, generating the initial control point of the target valve;

[0122] establishing a one-to-one mapping between the source side working condition cluster and the target side working condition cluster according to four types of characteristics of valve opening, valve differential pressure, medium temperature and medium physical property, obtaining the working condition cluster set of the target valve, and the one-to-one mapping between the source side working condition cluster and the target side working condition cluster is specifically:

[0123] clustering the source valve working condition data according to four types of characteristics of valve opening, valve differential pressure, medium temperature and medium physical property to form a source side working condition cluster set;

[0124] clustering the target valve working condition data based on the same four types of characteristics to form a target side working condition cluster set;

[0125] comparing the statistical characteristics of the source side and target side working condition clusters one by one, and establishing a one-to-one mapping relationship between the source side working condition cluster and the target side working condition cluster according to the minimum difference principle of characteristic mean value and distribution form;

[0126] According to the quantile retention rule, the source side residual characteristic baseline and the quantile threshold are converted into the residual characteristic baseline and the quantile threshold of the target valve, and the frequency band energy, the standard deviation and the spectral entropy baseline related to the residual amplification factor are updated synchronously, and the quantile retention rule is specifically:

[0127] Statistical analysis of the historical distribution of the source side residual characteristic baseline, extraction of a plurality of preset quantile points, including median, upper and lower quantile points and extreme quantile points;

[0128] On the distribution of the target valve working condition data, find the same value as the probability position of the corresponding quantile point on the source side, as the residual characteristic baseline quantile value of the target valve;

[0129] According to the principle of keeping the same distribution form, the source side quantile point and the target side quantile point are one-to-one corresponding mapping, completing the migration of the residual characteristic baseline and the quantile threshold, and ensuring that the threshold judgment of the target valve is consistent with the source valve in a statistical sense;

[0130] Collecting the working condition data of the target valve within the adaptation period, using the initial control point and the quantile threshold of the target valve to calculate the predicted flow and the residual, and combining the residual characteristics and the abnormal score to perform conformal calibration, obtaining the calibrated control point and the calibrated threshold of the target valve;

[0131] The calibration control points and calibration threshold of the target valve are applied to the playback data at the edge computing node for cross verification, and the improved results of the window residual mean and the window residual variance are compared. When the preset judgment condition is met, the online adaptation is completed, and when it is not met, it is rolled back and the time and reason are recorded, wherein the playback data refers to the working condition data collected and stored during the target valve adaptation period.

[0132] The present application realizes the rapid adaptation of new valves and new working conditions by introducing the order-preserving parameter deformation and working condition cluster isomorphism mapping mechanism. In the opening-flow characteristic aspect, the order-preserving deformation method with limited amplitude is adopted, and only the control points in the effective working interval of the target valve are adjusted to ensure the monotonicity and smoothness of the curve, avoiding the instability caused by large-scale parameter fitting in the traditional method. In the working condition division aspect, the corresponding relationship between the source side and the target side working condition clusters is established through the clustering mapping of the valve opening, differential pressure, medium temperature and physical property characteristics, and the residual baseline and quantile threshold are migrated combined with the quantile number preservation rule to ensure the consistency and comparability of the statistical characteristics. The conformal calibration and playback verification mechanism is introduced in the adaptation process, and the calibration results are verified offline using historical data. Only when the residual mean and variance are improved, the online adaptation is completed, which improves the robustness and reliability of parameter migration.

[0133] Embodiment 1

[0134] In order to verify the feasibility of the present application in implementation, the present application is applied to the main steam pipeline of a large thermal power plant, and a steam regulating valve is installed for regulating the steam flow from the boiler to the steam turbine. The steam medium temperature fluctuates between 180℃ and 220℃, and the pressure before the valve is between 1.6MPa and 2.0MPa. The valve has been running for more than a year, and the original prediction method based on the factory opening-flow characteristic curve has decreased in accuracy, with a deviation of more than 6%. When running at a small opening (below 20%), the valve has a slight leakage due to wear of the sealing pair, but the original judgment method based on a fixed threshold value cannot identify it.

[0135] After applying the method of the present application, first, the valve opening, pre-valve pressure, post-valve pressure, differential pressure, medium temperature, actuator current, pipeline vibration signal and acoustic signal are collected by sensors installed on the valve actuator and pipeline, after time alignment and noise filtering, a working condition data set is formed. Based on the data set, a digital twin model is constructed, wherein the control points of the factory opening-flow curve are used as the baseline for the physical calculation unit, combined with the steam density and viscosity to correct the differential pressure-flow relationship; the correction amount is output by the data-driven model using the actuator current, vibration and acoustic signals in the historical window as input, and the physical predicted flow is superimposed to form the fusion predicted flow. Then, the predicted flow is compared with the measured flow of the flowmeter to obtain the residual sequence.

[0136] At the 90th day of continuous operation, the system detected that the residual variance exceeded the preset threshold, triggering a self-calibration update that automatically corrected the opening-flow control point. After calibration, the prediction error decreased from 5.8% to 2.2%. At the 120th day, the residual standard deviation increased to 0.27 when the valve was operating at a small opening, and the system amplified the weak anomaly through the residual amplification factor mechanism, achieving a similarity of 0.83 with the historical leakage template, successfully triggering an early warning. In contrast, the traditional threshold method did not produce an alarm.

[0137] In subsequent applications, the method was also adapted to another steam valve. Through the order-preserving parameter deformation mechanism, the control points of the original valve were mapped to the new valve, and the residual feature baseline and quantile threshold were generated by combining the isomorphic mapping of the operating condition cluster. Within the 48-hour adaptation period, the prediction error after calibration decreased from 7.0% to 2.5%, and after verification by replaying the data, the adaptation was successful and went online.

[0138] Table 1 Valve flow monitoring and anomaly detection results

[0139] Time (days) Actual flow (t / h) Original method prediction error (%) Invention prediction error (%) Residual standard deviation Self-calibration Early warning trigger 30 105.2 3.2 1.6 0.10 No No 60 104.7 4.5 1.9 0.14 No No 90 103.9 5.8 2.2 0.21 Yes No 120 102.5 6.0 2.4 0.27 No Early warning (leak) 150 101.8 6.3 2.5 0.29 No Alarm (leak)

[0140] As can be seen from the data in Table 1, within the first 60 days of valve operation, the error between the valve prediction results and the actual measured flow gradually accumulated, increasing from 3.2% at the 30th day to 4.5% at the 60th day. Although the error increased, it was still within an acceptable range, with the residual standard deviation remaining between 0.10 and 0.14, indicating that the valve was operating in a relatively stable condition during this period, and the system did not trigger self-calibration and anomaly early warning.

[0141] At the 90th day, the prediction error further increased to 5.8%, and the residual standard deviation also rose to 0.21, exceeding the preset threshold. At this time, the system automatically triggered the self-calibration update mechanism, which corrected the opening-flow characteristic control points in the digital twin model. The prediction error of the method of the present application decreased to 2.2%, and the residual fluctuation was suppressed, demonstrating the role of the self-calibration mechanism in maintaining prediction accuracy over a long period of operation. In contrast, the original method did not update dynamically, and the error continued to increase, verifying the advantages of the present application.

[0142] At the 120th and 150th days, the valve entered small opening operating conditions, and the signs of leakage gradually became apparent. At this time, the residual standard deviation increased to 0.27 and 0.29, respectively, and the system amplified the weak anomaly through the residual amplification factor mechanism and combined with the quantile threshold of the operating condition cluster to determine. The results showed that the system triggered a leakage warning at the 120th day and further upgraded to a leakage alarm at the 150th day. The present application can identify small leaks that are difficult to detect by traditional methods, and has higher sensitivity and anomaly diagnosis capability.

[0143] The method can keep the prediction accuracy stable through the self-calibration mechanism in long-term operation, and can realize timely identification of weak abnormalities by using residual amplification and quantile threshold determination, thereby providing a more reliable valve flow monitoring method for an industrial site.

[0144] The above merely describes the preferred embodiments of the present application, but the protection scope of the present application is not limited thereto, and any person skilled in the art can make equivalent replacements or changes to the technical solutions and the inventive concept of the present application within the technical scope disclosed by the present application, which should be covered within the protection scope of the present application.

Claims

1. A method for valve real-time flow monitoring based on big data analysis, characterized in that, The method comprises the following steps: Collecting valve flow data, preprocessing the valve flow data to form a working condition data set; Based on the valve characteristic parameters and medium physical property parameters in the working condition data set, a digital twin model including a physical calculation unit and a correction calculation unit is constructed, the valve predicted flow is output by using the digital twin model, and the residual between the actual flow and the predicted flow is calculated to obtain a residual sequence; The residual sequence is statistically analyzed by a sliding time window, and it is determined whether the residual mean and variance exceed the preset threshold. When the over-limit condition is met, the update of the digital twin model parameters is performed under the steady state working condition, and the updated valve predicted flow and residual sequence are output; The updated residual sequence is extracted in time domain and frequency domain, and compared with the three types of feature vectors in the historical abnormal template library, i.e. leakage, blockage and hysteresis, to generate a residual amplification factor. The abnormal score is obtained by enhancing the residual sequence with the residual amplification factor; The abnormal score is compared with the quantile threshold of the corresponding working condition cluster, and the warning information or alarm information is output according to the comparison result, and the abnormal type label is marked; Based on the working condition data set, valve characteristic parameters, residual features and abnormal scores, a sequential parameter transformation and working condition cluster isomorphic mapping adaptation mechanism is executed on a big data platform. The monotonic spline transformation is performed on the source valve's opening-flow characteristic curve and residual feature distribution, the isomorphic mapping is performed on the working condition cluster with quantile preservation, the conformal calibration is performed on the valve predicted flow and residual sequence, and the transformation parameters and thresholds are issued after verification on the edge node with a playback set. If the verification is passed, the new valve or new working condition is adapted online; if the verification is not passed, the original parameters are kept and the rollback is recorded. The sequential parameter transformation and working condition cluster isomorphic mapping adaptation mechanism based on the working condition data set, valve characteristic parameters, residual features and abnormal scores, and executed on a big data platform, comprises: Determine the source valve and the target valve, read the working condition data set, the opening-flow characteristic control point, the working condition cluster set, the residual feature baseline, and the warning threshold and alarm threshold; Perform sequential parameter transformation on the opening-flow characteristic control point, limit the transformation within the target valve working opening range and keep it monotonic and smooth, limit the single transformation amplitude, and generate the initial control point of the target valve; According to the valve opening, valve differential pressure, medium temperature and medium physical property, a one-to-one mapping between the source side working condition cluster and the target side working condition cluster is established, and the working condition cluster set of the target valve is obtained; According to the quantile preservation rule, the source side residual feature baseline and quantile threshold are converted into the residual feature baseline and quantile threshold of the target valve, and the frequency band energy, standard deviation and spectral entropy baseline related to the residual amplification factor are updated synchronously; During the adaptation period, the working condition data of the target valve is collected, the initial control point and quantile threshold of the target valve are used to calculate the predicted flow and residual, and the conformal calibration is performed combined with the residual feature and abnormal score to obtain the calibration control point and calibration threshold of the target valve. The calibration control points and calibration threshold of the target valve are applied to the playback data at the edge computing node, and the improved results of the window residual mean and the window residual variance are compared. When the preset judgment condition is met, the online adaptation is completed, otherwise, the rollback is performed and the time and reason are recorded.

2. A valve real-time flow monitoring method based on big data analysis according to claim 1, characterized in that, The valve flow data includes valve opening data, valve inlet and outlet pressure data, valve differential pressure data, medium temperature data, actuator current data, vibration signal data, acoustic signal data, and flow data measured by a flowmeter.

3. A valve real-time flow monitoring method based on big data analysis according to claim 1, characterized in that, The preprocessing of the valve flow data includes time alignment processing, noise filtering processing, abnormal point elimination processing, and missing data completion processing on the collected valve flow data.

4. The real-time flow monitoring method of valves based on big data analysis as claimed in claim 1, wherein, The working condition cluster refers to clustering the valve flow data according to four key features of valve opening, valve differential pressure, medium temperature and medium physical property under the condition of fixed valve model and medium category, forming a stable data set with boundary constraints, and each working condition cluster is represented by feature boundary and unique identifier.

5. A valve real-time flow monitoring method based on big data analysis according to claim 1, characterized in that, The residual sequence is obtained by: reading the data in the working condition data set, constructing a digital twin model and instantiating it in an edge computing node, the digital twin model is composed of a physical calculation unit and a correction calculation unit, and the data in the working condition data set is input, and it runs at a uniform sampling period; constructing a physical calculation unit, establishing a monotonic spline mapping based on the factory opening-flow characteristic curve, parameterizing the factory opening-flow characteristic curve with several opening-flow characteristic control points, introducing the correction of medium physical property on differential pressure-flow relationship, and applying monotonicity and smoothness constraints, outputting a time-aligned physical predicted flow sequence; constructing a correction calculation unit, selecting the time series features of the working condition data within a preset historical window length as input data, using a fixed-layer feedforward neural network as the data-driven structure, normalizing the input data, and applying amplitude limiting constraints to the output results, generating a time-aligned correction amount sequence and corresponding confidence indication information; adding and fusing the physical predicted flow sequence and the correction amount sequence to generate a valve predicted flow sequence, performing physical feasibility verification and boundary truncation, and generating prediction uncertainty indication information according to physical constraints and confidence indication information; determining the actual flow sequence, when the flowmeter is available on site, using the verified flowmeter measurement value as the actual flow sequence, when the flowmeter is not available on site, performing differential pressure method flow calculation based on the valve, differential pressure and medium physical property to form the actual flow sequence, aligning the actual flow sequence and the valve predicted flow sequence, and then subtracting them by time to obtain the residual sequence and output.

6. A valve real-time flow monitoring method based on big data analysis according to claim 1, characterized in that, The output of the updated valve predicted flow and residual sequence includes: setting the sliding time window length and step, dividing the residual sequence into windows in time sequence, respectively calculating the residual mean and residual variance in each sliding time window; setting the mean threshold and variance threshold for each sliding time window, and determining the steady state condition in the same sliding time window, the steady state condition includes that the valve opening rate of change, the differential pressure rate of change and the medium temperature fluctuation of the multi-source working condition data do not exceed the pre-defined value; When the residual mean or residual variance of any sliding time window exceeds the corresponding threshold and the steady-state condition is met, a double-evidence trigger for digital twin model parameter update is initiated, including residual statistical over-limit evidence and multi-source working condition data consistency evidence; After the double-evidence trigger, a local order-preserving deformation update is performed on the physical calculation unit, with the discrete control points of the opening-flow characteristic curve as the update objects, and the deformation is only performed within the opening interval corresponding to the trigger sliding time window, maintaining the monotonicity and smoothness of the curve, and limiting the single deformation amplitude to not exceed the preset proportion; A restricted fine-tuning is performed on the correction calculation unit, adjusting the output-side parameters, using a segmented weighting strategy based on residual amplitude to suppress the influence of abnormal samples, setting the maximum step size and output range constraints, and generating an updated valve predicted flow sequence; A shadow twin is constructed for comparison and verification, and the historical data segment adjacent to the trigger sliding time window is selected for rapid verification, including two indicators of residual mean reduction and residual variance reduction. If the verification is passed, the update is written to the effective version and the updated valve predicted flow and corresponding residual sequence are output.

7. A valve real-time flow monitoring method based on big data analysis according to claim 1, characterized in that, The generation of the residual amplification factor, which enhances the residual sequence after the residual amplification factor to obtain an abnormal score, includes: Receiving the updated residual sequence, setting the length and step size of the sliding window, and dividing the residual sequence into consecutive analysis windows in chronological order; Calculating the time-domain statistics in each analysis window, including the mean, standard deviation, skewness, and kurtosis of the residual sequence, to form a time-domain feature vector; Performing spectral analysis on the residual sequence in each analysis window to extract low, medium, and high frequency energy, main peak frequency, and spectral entropy, forming a frequency-domain feature vector; Calculating the similarity between the time-domain feature vector and the frequency-domain feature vector and the feature vectors corresponding to the leakage template, blockage template, and hysteresis template in the historical abnormal template library to obtain similarity indicators for each abnormal type, and selecting the maximum value as the maximum similarity of the current analysis window; Normalizing the standard deviation, selected frequency band energy, and spectral entropy according to the corresponding working condition cluster historical baseline, and combining them with the maximum similarity by a preset weight, limiting the combination result within a preset range, and generating a residual amplification factor; Based on the baseline normalized mean of the residual amplitude in the analysis window, multiplying it by the residual amplification factor to obtain an abnormal score sequence.

8. A valve real-time flow monitoring method based on big data analysis according to claim 1, characterized in that, The comparison of the abnormal score with the quantile threshold of the corresponding working condition cluster, according to the comparison result, outputs the pre-warning information or alarm information, and labels the abnormal type tag, including: Determine the working condition cluster according to the working condition characteristics, retrieve the historical abnormal score distribution and historical residual amplification factor distribution of the working condition cluster as the reference data for the current window determination; Calculate two preset quantile thresholds on the historical abnormal score distribution, which are used as the pre-warning threshold and alarm threshold respectively; Compare the abnormal score of the current window with the pre-warning threshold and alarm threshold to obtain one of the three determination levels: normal, pre-warning, or alarm. The number of pre-warning duration windows and the number of alarm duration windows and the back-off margin are set, the determination levels of the continuous windows are accumulated and counted, when the accumulation reaches the corresponding duration window number, a pre-warning event or an alarm event is generated respectively, when the abnormal score of the continuous window is lower than the pre-warning threshold minus the back-off margin, the generated pre-warning or alarm state is cleared; The three types of similarity indexes of leakage, blockage and delay are read, the abnormal type with the highest similarity value is selected as the abnormal type label of the current window, and is associated with the abnormal score, the residual amplification factor and the working condition cluster identification of the current window; An event record is generated, the event record is written into the edge side event queue and is synchronized to the cloud end log, and the event count and the latest occurrence time of the corresponding working condition cluster are updated.

Images