Industrial cooler double bypass valve fault diagnosis early warning method and system

Abstract

The present application relates to the technical field of industrial cooler fault diagnosis, and discloses a method and system for diagnosing and early warning of faults of a double bypass valve of an industrial cooler. The method constructs a state feature vector by acquiring real-time operating parameters, performs dynamic time series analysis on the state feature vector to calculate sliding entropy values and Lyapunov exponents, determines state evolution trend characteristics, calculates a fault feature index in combination with a valve opening and a control signal deviation value, determines a warning threshold value based on a distribution variance of historical data, and determines a fault type and outputs warning information when the index exceeds the threshold value. The present application can predict potential faults of the bypass valve in advance, reduce the false positive rate, and improve the diagnostic accuracy.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of industrial cooler fault diagnosis, and particularly relates to an industrial cooler double bypass valve fault diagnosis and early warning method and system. BACKGROUND

[0002] Industrial coolers are key equipment widely used in modern industrial production. The double bypass valve, as an important control component, plays an irreplaceable role in maintaining system stable operation and adjusting cooling efficiency. With the continuous improvement of industrial automation, the reliability and safety of the cooling system are increasingly stringent. The double bypass valve adjusts the flow and pressure of the cooling medium to achieve precise control of the temperature of industrial equipment. However, it is prone to performance degradation or even failure due to mechanical wear, medium corrosion, control system failure, and other factors during long-term operation, which may cause system paralysis or safety accidents.

[0003] Current industrial cooler double bypass valve fault diagnosis technology mainly relies on periodic manual inspection, simple threshold monitoring, or post-analysis. These methods have obvious shortcomings. Traditional fault diagnosis methods lack dynamic time series analysis capability for valve state, and cannot capture early subtle changes in valve performance degradation, resulting in delayed fault warning and missed optimal maintenance opportunities. Existing technologies usually use single parameter or fixed threshold judgment methods, and fail to consider multi-dimensional feature information such as valve opening, control signal deviation, and system entropy, making it difficult to accurately identify various fault modes under complex working conditions, resulting in frequent false positives and false negatives. Existing methods cannot adapt to the dynamic characteristics of industrial cooling systems under different working conditions, lack adaptive early warning mechanisms, and cannot automatically adjust fault judgment standards according to actual operating environments, limiting the accuracy and practicality of fault diagnosis. SUMMARY

[0004] The industrial cooler double bypass valve fault diagnosis and early warning method and system provided by the embodiments of the present application can at least solve some of the problems existing in the prior art.

[0005] In a first aspect, the present application provides an industrial cooler double bypass valve fault diagnosis and early warning method, comprising:

[0006] Obtaining real-time operating parameters of the industrial cooler double bypass valve, and constructing an operating state feature vector of the industrial cooler double bypass valve based on the real-time operating parameters;

[0007] Performing dynamic time series analysis on the operating state feature vector, calculating the sliding entropy value of the operating state feature vector and the Lyapunov exponent of the reconstructed trajectory in the phase space, and determining the state evolution trend feature of the industrial cooler double bypass valve according to the sliding entropy value and the Lyapunov exponent;

[0008] According to the state evolution trend feature and the deviation value of the valve opening degree and the input control signal in the operating state feature vector, a fault feature index of the industrial cooler double bypass valve is calculated;

[0009] Based on the historical data distribution variance of the fault feature index, combined with the operating condition parameters of the industrial cooler, a fault warning threshold of the fault feature index is determined;

[0010] When the fault feature index exceeds the fault warning threshold, according to the feature combination of the fault feature index exceeding the threshold, the fault type of the industrial cooler double bypass valve is determined, and the corresponding fault warning information is output.

[0011] Obtain the real-time operating parameters of the industrial cooler double bypass valve, and construct an operating state feature vector of the industrial cooler double bypass valve based on the real-time operating parameters, including:

[0012] Obtain the real-time operating parameters of the first bypass valve and the second bypass valve of the industrial cooler double bypass valve during operation;

[0013] Perform time domain analysis on the real-time operating parameters, extract time domain feature parameters of the first bypass valve and the second bypass valve, and perform wavelet decomposition on the time domain feature parameters to obtain high frequency components and low frequency components of the first bypass valve and the second bypass valve;

[0014] According to the high frequency components, extract transient response features, according to the low frequency components, extract steady state operating features, combine the transient response features and the steady state operating features, and obtain dynamic and static feature parameters of the double bypass valve;

[0015] According to the dynamic and static feature parameters, construct a state space matrix of the first bypass valve and the second bypass valve, and determine a principal feature vector based on singular value decomposition results of the state space matrix;

[0016] Project the principal feature vector to a preset feature space to obtain a feature mapping matrix, and extract the operating state feature vector according to the feature mapping matrix.

[0017] Perform dynamic time series analysis on the operating state feature vector, calculate the sliding entropy value of the operating state feature vector and the Lyapunov index of the phase space reconstruction trajectory, and determine the state evolution trend feature of the industrial cooler double bypass valve according to the sliding entropy value and the Lyapunov index, including:

[0018] Divide the operating state feature vector into multiple time windows according to time sequence, and calculate the sliding entropy value of the operating state feature vector in each time window;

[0019] construct a phase space reconstructed trajectory based on the operating state feature vector, determine embedding dimension and time delay parameter of the phase space reconstructed trajectory;

[0020] reconstruct a phase space trajectory of the operating state feature vector according to the embedding dimension and the time delay parameter, and calculate a Lyapunov exponent of the phase space trajectory;

[0021] perform multi-scale fusion analysis on the sliding entropy value and the Lyapunov exponent, and construct a cooperative evolution matrix of the state feature of the double bypass valve;

[0022] perform feature decomposition on the cooperative evolution matrix to obtain a feature value sequence, sort the feature value sequence according to amplitude to obtain a dominant feature value corresponding to the largest amplitude, and construct a state evolution feature space based on the dominant feature value;

[0023] calculate a convergence radius and a divergence coefficient of a state trajectory in the state evolution feature space, and determine a state evolution trend feature of the double bypass valve of the industrial cooler according to the convergence radius and the divergence coefficient.

[0024] perform multi-scale fusion analysis on the sliding entropy value and the Lyapunov exponent, and construct a cooperative evolution matrix of the state feature of the double bypass valve, including:

[0025] construct a bivariate cooperative matrix according to the sliding entropy value and the Lyapunov exponent, perform recursive quantization analysis on the bivariate cooperative matrix, and establish a dynamic correlation network of the state feature;

[0026] extract topological features and transmission characteristics of the dynamic correlation network, and determine a coupling relationship between state features according to the topological features and the transmission characteristics;

[0027] perform wavelet transform on the coupling relationship to obtain wavelet coefficients at different scales, and divide the wavelet coefficients into a high-frequency subband and a low-frequency subband according to an energy distribution rule;

[0028] construct a multi-scale feature matrix based on the high-frequency subband and the low-frequency subband, and perform singular value decomposition on the multi-scale feature matrix to obtain a principal component feature vector;

[0029] project the principal component feature vector to a preset state evolution space to obtain a state evolution projection matrix, and construct a cooperative evolution matrix of the state feature of the double bypass valve according to the state evolution projection matrix.

[0030] calculate a fault feature index of the double bypass valve of the industrial cooler according to the state evolution trend feature and a deviation value of a valve opening degree and an input control signal in the operating state feature vector, including:

[0031] constructing a dynamic transfer matrix based on the state evolution trend characteristics, and constructing a valve response feature matrix based on a deviation value of a valve opening in the operating state feature vector and an input control signal;

[0032] performing collaborative mapping of the dynamic transfer matrix and the valve response feature matrix to obtain a dynamic-static feature transfer path and a valve response topology correlation degree of the double bypass valve;

[0033] extracting valve opening distribution characteristics of the double bypass valve under different operating states according to the dynamic-static feature transfer path, and constructing a spatial mapping matrix of the valve opening;

[0034] calculating an opening response correlation strength and an opening synchronicity index of the double bypass valve based on the valve response topology correlation degree, and constructing a correlation mapping matrix of the valve opening;

[0035] performing feature fusion of the spatial mapping matrix of the valve opening and the correlation mapping matrix of the valve opening to obtain a valve opening comprehensive evaluation index of the double bypass valve, and calculating a fault feature index of the double bypass valve of the industrial cooler according to the valve opening comprehensive evaluation index.

[0036] determining a fault warning threshold of the fault feature index based on a historical data distribution variance of the fault feature index and in combination with an operating condition parameter of the industrial cooler, including:

[0037] performing distribution characteristic analysis on the fault feature index to construct a historical distribution feature vector of the fault feature index;

[0038] calculating a distribution skewness and a distribution kurtosis of the fault feature index according to the historical distribution feature vector to obtain a historical data distribution variance of the fault feature index;

[0039] classifying the operating condition parameter according to a condition type to construct a condition feature vector, and calculating a correlation coefficient and a contribution degree coefficient of each parameter in the condition feature vector;

[0040] constructing a condition parameter combination coefficient matrix based on the correlation coefficient and the contribution degree coefficient, and constructing a parameter distribution matrix based on the historical data distribution variance;

[0041] performing coupling mapping of the parameter distribution matrix and the condition parameter combination coefficient matrix to extract an association characteristic of the fault feature index and the condition parameter, constructing an adaptive threshold mapping function based on the association characteristic, and determining the fault warning threshold of the fault feature index according to the adaptive threshold mapping function.

[0042] The parameter distribution matrix is coupled and mapped with the working condition parameter combination coefficient matrix, the correlation characteristics of the fault characteristic index and the working condition parameters are extracted, the adaptive threshold mapping function is constructed based on the correlation characteristics, and the fault warning threshold of the fault characteristic index is determined according to the adaptive threshold mapping function, including:

[0043] The distribution density function and the cumulative distribution function of the fault characteristic index are calculated based on the parameter distribution matrix, and the probability distribution feature vector of the fault characteristic index is constructed;

[0044] The mutual information and the conditional entropy between the working condition parameters are calculated by using the working condition parameter combination coefficient matrix, and the state transition matrix of the working condition feature is constructed;

[0045] The probability distribution feature vector is mapped and operated with the state transition matrix, the time-varying correlation degree of the fault characteristic index and the working condition parameters is calculated based on the mutual information and the conditional entropy as the correlation characteristics;

[0046] The dynamic weight matrix of the working condition parameters is constructed according to the time-varying correlation degree, the comprehensive influence weight of the working condition parameters is calculated by using the dynamic weight matrix, and the multi-dimensional mapping space of the working condition feature is constructed based on the comprehensive influence weight;

[0047] The adaptive threshold mapping function considering the working condition conversion is established in the multi-dimensional mapping space, the reference threshold of the working condition steady state interval and the dynamic compensation factor of the working condition conversion interval in the adaptive threshold mapping function are adaptively fused, and the fault warning threshold of the fault characteristic index is determined.

[0048] The second aspect of the embodiment of the application provides an industrial cooler double bypass valve fault diagnosis and early warning system, including:

[0049] The first unit is used for acquiring real-time running parameters of the industrial cooler double bypass valve, and constructing a running state feature vector of the industrial cooler double bypass valve based on the real-time running parameters;

[0050] The second unit is used for performing dynamic time sequence analysis on the running state feature vector, calculating the sliding entropy value of the running state feature vector and the Lyapunov index of the phase space reconstruction trajectory, and determining the state evolution trend feature of the industrial cooler double bypass valve according to the sliding entropy value and the Lyapunov index;

[0051] The third unit is used for calculating the fault characteristic index of the industrial cooler double bypass valve according to the state evolution trend feature and the deviation value of the valve opening and the input control signal in the running state feature vector;

[0052] A fourth unit is configured to determine a fault early warning threshold of the fault feature index based on a variance of a historical data distribution of the fault feature index and in combination with an operating condition parameter of the industrial cooler;

[0053] A fifth unit is configured to determine a fault type of a double bypass valve of the industrial cooler according to a feature combination of the fault feature index exceeding the threshold when the fault feature index exceeds the fault early warning threshold, and output corresponding fault early warning information.

[0054] In a third aspect, an electronic device is provided, and the electronic device comprises:

[0055] a processor;

[0056] a memory for storing processor-executable instructions;

[0057] The processor is configured to invoke the instructions stored in the memory to execute the method described above.

[0058] In a fourth aspect, a computer-readable storage medium is provided, and the computer-readable storage medium stores computer program instructions, and the computer program instructions are executed by a processor to implement the method described above.

[0059] The industrial cooler double bypass valve fault diagnosis and early warning method provided by the application can accurately capture the slight change trend in the valve operation process by constructing an operating state feature vector through real-time operating parameters and performing dynamic time series analysis on the feature vector, early identification of faults is achieved, and the sensitivity and accuracy of fault early warning are effectively improved.

[0060] The method of the application determines the state evolution trend feature by calculating the sliding entropy value and the Lyapunov index, and forms a complete multi-dimensional fault feature evaluation system by combining the deviation value analysis of the valve opening and the control signal, which not only can detect conventional faults, but also can identify hidden faults that are difficult to find by traditional methods, greatly expanding the range of fault diagnosis.

[0061] The method of the application dynamically adjusts the fault early warning threshold based on the variance of the historical data distribution, and comprehensively judges in combination with the actual operating condition parameters of the industrial cooler, effectively reduces the false positive rate, and can accurately determine the fault type according to the feature combination of the fault feature index, provides clear fault positioning and processing suggestions for maintenance personnel, and significantly improves the operation reliability and maintenance efficiency of the industrial cooling system. BRIEF DESCRIPTION OF DRAWINGS

[0062] Figure 1 FIG. 1 is a flowchart of an industrial cooler double bypass valve fault diagnosis and early warning method according to an embodiment of the application.

[0063] Figure 2A flowchart of the industrial cooler double bypass valve fault diagnosis and early warning method of the embodiment of the present application is shown in Figure 1. DETAILED DESCRIPTION

[0064] In an alternative embodiment, the embodiment of the present application provides an industrial cooler double bypass valve fault diagnosis and early warning method. Figure 1 A flowchart of the industrial cooler double bypass valve fault diagnosis and early warning method of the embodiment of the present application is shown in Figure 1. Figure 1 As shown in Figure 1, the method comprises the following steps.

[0065] Obtaining real-time operation parameters of the industrial cooler double bypass valve, and constructing an operation state feature vector of the industrial cooler double bypass valve based on the real-time operation parameters;

[0066] Performing dynamic time series analysis on the operation state feature vector, calculating a sliding entropy value of the operation state feature vector and a Lyapunov exponent of a reconstructed trajectory of a phase space, and determining a state evolution trend feature of the industrial cooler double bypass valve according to the sliding entropy value and the Lyapunov exponent;

[0067] According to the state evolution trend feature and a deviation value of a valve opening degree and an input control signal in the operation state feature vector, calculating a fault feature index of the industrial cooler double bypass valve;

[0068] Based on a historical data distribution variance of the fault feature index, combining an operation condition parameter of the industrial cooler, and determining a fault early warning threshold value of the fault feature index;

[0069] When the fault feature index exceeds the fault early warning threshold value, determining a fault type of the industrial cooler double bypass valve according to a feature combination of the fault feature index exceeding the threshold value, and outputting corresponding fault early warning information.

[0070] In an alternative embodiment, obtaining real-time operation parameters of the industrial cooler double bypass valve, and constructing an operation state feature vector of the industrial cooler double bypass valve based on the real-time operation parameters comprises the following steps.

[0071] Obtaining real-time operation parameters of the first bypass valve and the second bypass valve of the industrial cooler double bypass valve during operation;

[0072] Performing time domain analysis on the real-time operation parameters, extracting time domain feature parameters of the first bypass valve and the second bypass valve, performing wavelet decomposition on the time domain feature parameters, and obtaining high-frequency components and low-frequency components of the first bypass valve and the second bypass valve;

[0073] According to the high-frequency component, a transient response feature is extracted, according to the low-frequency component, a steady-state operation feature is extracted, the transient response feature and the steady-state operation feature are combined, and a dynamic and static feature parameter of the double bypass valve is obtained;

[0074] According to the dynamic and static feature parameter, a state space matrix of the first bypass valve and the second bypass valve is constructed, and a principal feature vector is determined based on a singular value decomposition result of the state space matrix;

[0075] The principal feature vector is projected to a preset feature space to obtain a feature mapping matrix, and the operation state feature vector is extracted according to the feature mapping matrix.

[0076] Figure 2 The flowchart of the industrial cooler double bypass valve fault diagnosis and early warning method of the embodiment of the application is shown in FIG. Figure 2 As shown in FIG.

[0077] In the industrial cooling system, the double bypass valve as a key control element, its operation state directly affects the stability and energy efficiency of the system. The double bypass valve in the embodiment includes a first bypass valve and a second bypass valve, which work together to achieve precise flow and pressure control.

[0078] The first step of constructing the feature vector is to obtain the real-time operation parameters of the first bypass valve and the second bypass valve of the industrial cooler double bypass valve during operation. Specifically, through a plurality of sensors installed on the valve body, data is collected, including but not limited to valve opening, flow, pressure, temperature and other parameters. These sensors are connected to the data acquisition system through the industrial field bus, and data is continuously collected at a sampling frequency of 100Hz. For example, the opening of the first bypass valve in a certain cooling system changes in the range of 0-100%, and the pressure difference during normal operation is 0.5-1.2MPa; the opening of the second bypass valve changes in the range of 0-80%, and the pressure difference during normal operation is 0.3-0.8MPa. These real-time operation parameters are stored in a time series database, providing a basis for subsequent analysis.

[0079] After obtaining the real-time operation parameters, time domain analysis is performed on the real-time operation parameters, and time domain feature parameters of the first bypass valve and the second bypass valve are extracted. Time domain analysis mainly investigates the change characteristics of the signal in the time dimension. For the operation parameters of the first bypass valve and the second bypass valve, the mean, standard deviation, kurtosis, skewness, maximum value, minimum value and other statistical characteristics are calculated. For example, for the pressure data of the first bypass valve collected within 10 minutes, the mean is 0.85MPa, the standard deviation is 0.12MPa, the kurtosis is 2.76, and the skewness is 0.31. These statistical characteristics can reflect the basic distribution characteristics and fluctuation of the parameters.

[0080] After the time domain feature parameter extraction is completed, the time domain feature parameters are wavelet decomposed to obtain high frequency components and low frequency components of the first bypass valve and the second bypass valve. The wavelet decomposition technology is used to process these parameters, and the signal is decomposed into high frequency components and low frequency components. In the wavelet decomposition process, db4 wavelet is selected as the mother wavelet, and the signal is decomposed for 5 levels. The high frequency components mainly contain the rapid change and transient characteristics of the signal, and the low frequency components reflect the slow change trend and steady state characteristics of the signal. In the specific operation, the collected time domain data is segmented according to 1024 points as a window, and the overlap rate between the windows is 50%. For example, after the wavelet decomposition of the flow data of the second bypass valve, the obtained high frequency components can reflect the mutation and fluctuation of the flow, and the low frequency components can reflect the stable change trend of the flow.

[0081] Based on the high frequency components and low frequency components obtained by wavelet decomposition, transient response features are extracted according to the high frequency components, and steady state operation features are extracted according to the low frequency components. The transient response features mainly focus on the response characteristics of the valve when starting, closing or load mutation. By analyzing the amplitude, energy distribution and change rate of the high frequency components, the characteristic parameters of the transient response can be obtained. For example, the energy proportion of the high frequency component can reflect the dynamic response strength of the system, and the zero-crossing rate of the high frequency component can reflect the oscillation frequency of the system. For the first bypass valve, in the starting stage, the high frequency energy proportion is usually 25%-35% of the total energy, and the zero-crossing rate is 12-18 times per second. The steady state operation features are obtained by analyzing the trend, periodicity and smoothness of the low frequency components. For example, autocorrelation analysis of the low frequency component can identify the periodic characteristics of the system, and the coefficient of variation of the low frequency component can evaluate the stability of the steady state operation of the system. For the second bypass valve, in the normal operation state, the coefficient of variation of the low frequency component usually remains between 0.05-0.15.

[0082] The transient response features and the steady state operation features are combined to obtain the dynamic and static characteristic parameters of the double bypass valve. These parameters contain both the dynamic response information of the valve and the steady state operation information, and can comprehensively reflect the working state of the valve. The dynamic and static characteristic parameters usually contain 20-30 dimensions, for example: the dynamic characteristics include response time, overshoot, regulation time, oscillation decay ratio, etc.; the static characteristics include steady state error, stability index, linearity, hysteresis characteristics, etc.

[0083] According to the dynamic and static characteristic parameters, a state space matrix of the first bypass valve and the second bypass valve is constructed. The state space matrix is a mathematical model used to describe the dynamic behavior of a system. During the construction process, the dynamic and static characteristic parameters are organized into a matrix form, with each row representing the characteristics at a time point and each column representing different characteristic dimensions. For example, for data with a monitoring period of 24 hours and a sampling interval of 1 hour, the constructed state space matrix has a dimension of 24x25 (24 time points, 25 characteristic dimensions). The first bypass valve and the second bypass valve construct their respective state space matrices for subsequent independent analysis.

[0084] Based on the singular value decomposition results of the state space matrix, the principal characteristic vectors are determined. Singular value decomposition can decompose the original matrix into the product of three matrices, which contains the main characteristic information of the original matrix. By calculating the size of the singular value, the number of principal characteristic vectors can be determined. Generally, the characteristic vectors corresponding to the first few singular values with a cumulative contribution rate of 95% are selected as the principal characteristic vectors. For example, singular value decomposition is performed on the state space matrix of the first bypass valve, and the first 5 singular values obtained are 45.6, 22.3, 10.8, 5.2, and 2.7, with a cumulative contribution rate of 96.3%. Therefore, the first 5 characteristic vectors are selected as the principal characteristic vectors.

[0085] The principal characteristic vectors are projected into a pre-set characteristic space to obtain a characteristic mapping matrix. The pre-set characteristic space is a standard space determined according to expert experience and historical data analysis, which defines the characteristic distribution under normal operating conditions. During the projection process, a linear mapping method is used to map the principal characteristic vectors to each dimension of the pre-set characteristic space. For example, if the pre-set characteristic space is 8-dimensional, the 5-dimensional principal characteristic vectors obtained earlier are mapped to 8-dimensional vectors through a conversion matrix. This mapping process can highlight important characteristic information while reducing the influence of noise and redundant information.

[0086] The operating state characteristic vector is extracted according to the characteristic mapping matrix. The characteristic mapping matrix contains the position information of the double bypass valve in the pre-set characteristic space. By combining and normalizing these information, the final characteristic vector is obtained. Specifically, the mean vector of the characteristic mapping matrix is calculated and normalized to obtain a characteristic vector that can represent the current operating state. For example, the mean vector of the characteristic mapping matrix of the second bypass valve is calculated as [0.72, 0.56, 0.83, 0.41, 0.65, 0.29, 0.77, 0.58], and after normalization, the operating state characteristic vector is obtained as [0.15, 0.11, 0.17, 0.08, 0.13, 0.06, 0.16, 0.12].

[0087] The operating state feature vector constructed in this way has the characteristics of moderate dimension and rich information, and can accurately reflect the operating state of the double bypass valve of the industrial cooler. The feature vector can be used in subsequent state monitoring, fault diagnosis and predictive maintenance application scenarios, providing important support for the safe and stable operation of the industrial cooling system.

[0088] In actual application, the system will continuously collect the operating parameters of the double bypass valve and update the operating state feature vector in real time according to the above method. By comparing the differences between the current feature vector and the historical feature vector, abnormal changes in the valve operating state can be found in time, and appropriate maintenance measures can be taken. For example, after a certain industrial cooling system has been running continuously for 3 months, the feature vector of the first bypass valve deviates significantly in the 5th and 7th dimensions, from the normal values of 0.13 and 0.16 to 0.21 and 0.09, respectively. The system immediately issues a warning, and maintenance personnel find that the valve sealing surface is slightly worn, which is repaired in time to avoid system downtime due to valve failure.

[0089] Collect the valve operating data within 5 minutes, process the data in segments, and each segment is 10 seconds long; calculate the statistical features of each segment of data, including mean, standard deviation, peak factor, waveform factor, etc.; average the features of each segment to obtain the final time domain feature parameters. For the first bypass valve, the waveform factor under normal state is about 1.2-1.4, and the peak factor is about 2.8-3.2; for the second bypass valve, the waveform factor under normal state is about 1.1-1.3, and the peak factor is about 2.6-3.0.

[0090] The time domain feature parameter sequence is preprocessed, including detrending and normalization; select an appropriate wavelet basis function, db4 wavelet is used in this embodiment; set the decomposition level to 5; perform wavelet decomposition algorithm to obtain 5 layers of detail coefficients and 1 layer of approximation coefficients; combine the 1-2 layer detail coefficients into high frequency components, and combine the 3-5 layer detail coefficients and approximation coefficients into low frequency components. The energy of the high frequency component usually accounts for 10%-30% of the total energy, and the energy of the low frequency component usually accounts for 70%-90% of the total energy.

[0091] The energy distribution of the high frequency component is calculated to obtain the energy concentration index; the amplitude statistical features of the high frequency component are calculated, including the maximum amplitude, the root mean square value, etc.; the time-frequency characteristics of the high frequency component are analyzed to extract the main frequency components; the change rate of the high frequency component is calculated, including the rise time, the fall time, etc. For the first bypass valve, the high frequency energy concentration under normal state is about 0.65-0.75, and the main frequency components are concentrated in 15-25Hz; for the second bypass valve, the high frequency energy concentration under normal state is about 0.60-0.70, and the main frequency components are concentrated in 12-22Hz.

[0092] Trend characteristics of low-frequency components are calculated, including slope, curvature, etc.; periodicity of low-frequency components is analyzed, and main period length is extracted; stationarity indicators of low-frequency components are calculated, including coefficient of variation, steady-state error, etc.; correlation characteristics of low-frequency components are analyzed, and autocorrelation coefficients are calculated. For the first bypass valve, the low-frequency trend slope under normal state is close to 0, and the coefficient of variation is about 0.08-0.12; for the second bypass valve, the low-frequency trend slope under normal state is close to 0, and the coefficient of variation is about 0.06-0.10.

[0093] The transient response characteristics and steady-state operation characteristics are normalized; the weight coefficients of each characteristic can be determined based on expert experience or historical data analysis; the characteristics are combined by weighting according to the weight coefficients; the combined results are dimensionally reduced to retain the main information. The final dynamic and static characteristic parameter vector dimension is usually 20-30 dimensions, of which the transient characteristics and steady-state characteristics each account for about 50% of the proportion.

[0094] The dynamic and static characteristic parameters for 24 consecutive hours are collected, with a sampling interval of 1 hour; the characteristic parameters at each time point are organized as a row vector; the row vectors of all time points are arranged in chronological order to form a state space matrix; the matrix is preprocessed, including normalization and singularity removal. For the first bypass valve and the second bypass valve, their respective state space matrices are constructed, and the matrix dimensions are both 24x25.

[0095] The singular value decomposition algorithm is performed on the state space matrix; the contribution rate and cumulative contribution rate of each singular value are calculated; the number of principal characteristic vectors is determined according to the cumulative contribution rate threshold (usually 95%); the corresponding left singular vectors are extracted as the principal characteristic vectors. For the first bypass valve, usually the first 5-7 characteristic vectors are selected; for the second bypass valve, usually the first 4-6 characteristic vectors are selected.

[0096] A preset feature space is constructed, usually defined based on historical data and expert knowledge; a mapping matrix is designed to map the principal characteristic vector space to the preset feature space; a linear mapping operation is performed to obtain a feature mapping matrix; the mapping results are post-processed, including normalization and outlier detection. The preset feature space is usually 8-12 dimensions, which can cover the main operating state characteristics of the valve.

[0097] The mean vector of the feature mapping matrix is calculated; the mean vector is normalized; the confidence interval of the vector is calculated for anomaly detection; the current operating state is evaluated by comparing with the historical characteristic vectors. The final operating state characteristic vector dimension is usually 8-12 dimensions, each dimension representing an aspect of the valve operating state, such as response speed, stability, linearity, etc.

[0098] In an alternative embodiment, the operating state feature vector is subjected to dynamic time series analysis, the sliding entropy value of the operating state feature vector and the Lyapunov exponent of the phase space reconstruction trajectory are calculated, and the state evolution trend feature of the industrial cooler double bypass valve is determined according to the sliding entropy value and the Lyapunov exponent, including:

[0099] The operating state feature vector is divided into a plurality of time windows according to the time sequence, and the sliding entropy value of the operating state feature vector is calculated in each time window;

[0100] A phase space reconstruction trajectory is constructed based on the operating state feature vector, and the embedding dimension and time delay parameter of the phase space reconstruction trajectory are determined;

[0101] The phase space trajectory of the operating state feature vector is reconstructed according to the embedding dimension and the time delay parameter, and the Lyapunov exponent of the phase space trajectory is calculated;

[0102] The sliding entropy value and the Lyapunov exponent are subjected to multi-scale fusion analysis, and a cooperative evolution matrix of the double bypass valve state feature is constructed;

[0103] The cooperative evolution matrix is subjected to feature decomposition to obtain a feature value sequence, the feature value sequence is sorted according to the amplitude to obtain a dominant feature value corresponding to the largest amplitude, and a state evolution feature space is constructed based on the dominant feature value;

[0104] The convergence radius and divergence coefficient of the state trajectory are calculated in the state evolution feature space, and the state evolution trend feature of the industrial cooler double bypass valve is determined according to the convergence radius and the divergence coefficient.

[0105] The operating state feature vector is subjected to dynamic time series analysis, the sliding entropy value of the operating state feature vector and the Lyapunov exponent of the phase space reconstruction trajectory are calculated, and the state evolution trend feature of the industrial cooler double bypass valve is determined according to the sliding entropy value and the Lyapunov exponent. The operating state feature vector usually contains 8-12 dimensions, and each dimension represents a specific aspect of the valve state.

[0106] The running state feature vector is divided into multiple time windows according to a time sequence, and a sliding entropy value of the running state feature vector is calculated in each time window. The division of the time window adopts a sliding window technology, the window length is 120 minutes, the sliding step is 10 minutes, and the adjacent window overlap rate is 91.7%. For the feature vector sequence in each time window, the sample entropy value is calculated, and the sample entropy can measure the complexity and irregularity of the time sequence. The specific calculation process includes: determining the template length m=2 and the similarity tolerance r=0.2*standard deviation; counting the number of template matches; calculating the conditional probability; taking the negative natural logarithm to obtain the sample entropy value. For example, for the first bypass valve in the normal running state, the typical sliding entropy value range is 0.45-0.65; for the first bypass valve with slight wear, the sliding entropy value will rise to 0.75-0.95, indicating that the system complexity increases.

[0107] A phase space reconstruction trajectory is constructed based on the running state feature vector, and an embedding dimension and a time delay parameter of the phase space reconstruction trajectory are determined. Phase space reconstruction is an effective means to reveal the dynamic characteristics of the system, and based on the Takens embedding theorem, a multi-dimensional phase space can be reconstructed from a single variable time sequence. Determining appropriate embedding dimension and time delay parameter is crucial for accurate reconstruction of the phase space. The time delay parameter determination adopts the mutual information method, and the mutual information between the time sequence and its delayed sequence is calculated, and the first local minimum point is taken as the time delay parameter. For the opening data of the second bypass valve, the typical time delay parameter is 5-8 sampling points. The embedding dimension determination adopts the false nearest neighbor method, and the false nearest neighbor ratio under different embedding dimensions is calculated, and when the false nearest neighbor ratio is lower than 5%, the dimension is taken as the embedding dimension. For the pressure fluctuation data, the typical embedding dimension is 4-6.

[0108] A phase space trajectory of the running state feature vector is reconstructed according to the embedding dimension and the time delay parameter, and a Lyapunov exponent of the phase space trajectory is calculated. Lyapunov exponent is an important parameter for characterizing the degree of chaos of the system, and a positive Lyapunov exponent indicates that the system has chaotic characteristics, and the larger the exponent value, the stronger the instability of the system. In the process of reconstructing the phase space trajectory, each component of the running state feature vector is regarded as an independent variable, and the Lyapunov exponent is calculated by reconstructing and calculating respectively. The steps for calculating the Lyapunov exponent include: finding the nearest neighbor points in the phase space; tracking the evolution trajectory of the nearest neighbor points; calculating the logarithmic value of the trajectory divergence rate; linear fitting the divergence rate, and the slope is the maximum Lyapunov exponent. For the stable running double bypass valve, the maximum Lyapunov exponent is usually 0.01-0.05; when the valve has a slight fault, the maximum Lyapunov exponent will rise to 0.08-0.15, indicating that the stability of the system decreases.

[0109] The sliding entropy value and the Lyapunov exponent are subjected to multi-scale fusion analysis to construct a cooperative evolution matrix of the state characteristics of the double bypass valve. The multi-scale fusion analysis considers the differences in system behavior at different time scales and can more comprehensively depict the dynamic characteristics of the system. The cooperative evolution matrix describes the change relationship of the sliding entropy value and the Lyapunov exponent at multiple time scales. The steps of constructing the cooperative evolution matrix include: multi-scale decomposition of the sliding entropy value and the Lyapunov exponent sequence, the scale factor ranges from 1 to 20; calculation of the entropy value and the Lyapunov exponent at each scale; construction of a two-dimensional matrix, the row representing different scales and the column representing time sequences; normalization processing of the matrix. For example, the cooperative evolution matrix of the entropy value and the Lyapunov exponent of the first bypass valve of an industrial cooling system at different scales is a 20x48 matrix (20 scales, 48 time points), and the matrix element value ranges from 0 to 1, representing the normalized entropy value and Lyapunov exponent.

[0110] The cooperative evolution matrix is subjected to feature decomposition to obtain a feature value sequence, the feature value sequence is sorted according to the amplitude size to obtain a dominant feature value corresponding to the largest amplitude, and a state evolution characteristic space is constructed based on the dominant feature value. Feature decomposition can extract the main mode in the cooperative evolution matrix and simplify the representation of the system. The steps of feature decomposition include: calculating the covariance matrix of the cooperative evolution matrix; performing eigenvalue decomposition on the covariance matrix; sorting according to the eigenvalue size; selecting the first few eigenvalues with a cumulative contribution rate of 90% as the dominant eigenvalues. For example, the cooperative evolution matrix of the second bypass valve is subjected to feature decomposition, and the first three eigenvalues obtained are 4.82, 2.15 and 0.93, with a cumulative contribution rate of 91.7%, so these three eigenvalues are selected to construct the state evolution characteristic space. The state evolution characteristic space is a subspace spanned by the eigenvectors corresponding to the dominant eigenvalues, which can capture the main features of the system state evolution.

[0111] The convergence radius and divergence coefficient of the state trajectory are calculated in the state evolution characteristic space, and the state evolution trend characteristics of the industrial cooler double bypass valve are determined according to the convergence radius and the divergence coefficient. The state trajectory reflects the change path of the system state over time, and the convergence radius and the divergence coefficient describe the geometric characteristics and dynamic characteristics of the trajectory. The convergence radius calculation step includes: identifying an attractor of the state trajectory in the state evolution characteristic space; calculating the average distance of the trajectory point to the attractor; calculating the standard deviation of the distance; and defining the convergence radius as an interval formed by adding and subtracting three times the standard deviation from the average distance. The divergence coefficient calculation step includes: selecting adjacent trajectory point pairs; calculating the rate of change of the distance between the trajectory point pairs over time; statistically analyzing the rate of change; and defining the divergence coefficient as the average value of the rate of change. For a stably operating double bypass valve, the convergence radius is usually 0.05-0.15, and the divergence coefficient is 0.01-0.03; for a valve with a state tending to be unstable, the convergence radius increases to 0.2-0.4, and the divergence coefficient increases to 0.05-0.1.

[0112] The running state characteristic vector is divided into multiple time windows according to the time sequence, the window length is set to 120 minutes, corresponding to 720 sample points of the characteristic vector (when the sampling frequency is 0.1 Hz); the sliding step is set to 10 minutes, corresponding to 60 sample points; starting from the beginning of the time sequence, slide back 60 sample points each time, and intercept a time window with a length of 720 sample points; for 24 hours of monitoring data, 144 time windows can be obtained.

[0113] The sliding entropy value is calculated in each time window, the characteristic vector sequence in the time window is standardized to have a mean of 0 and a standard deviation of 1; the template length m=2 is set, i.e. taking 2 consecutive sample points as a template each time; the similarity tolerance r=0.2×standard deviation is set, i.e. r=0.2; the distance between all m-dimensional templates is calculated, and the number of template pairs with a distance less than r A is counted; the template length m+1=3 is set, and the above process is repeated, and the number of template pairs with a distance less than r B is counted; the sample entropy value is calculated as -ln(B / A). For example, for the characteristic vector sequence in a certain time window of the first bypass valve, A=385 and B=129 are counted, and the sample entropy value is -ln(129 / 385)=1.09.

[0114] The sliding entropy values of different feature vector components are calculated respectively, and then the average value is taken as the overall sliding entropy value. For example, for an 8-dimensional feature vector, the sliding entropy values of the 8 components calculated are 0.82, 0.76, 0.95, 0.88, 0.79, 0.91, 0.84, and 0.73, and the overall sliding entropy value is 0.835. The change trend of the sliding entropy value can reflect the evolution of system complexity. For example, when the sliding entropy value of the second bypass valve gradually increases from the initial 0.56 to 0.83, it indicates that the system complexity increases and an abnormal state occurs.

[0115] Based on the running state feature vector, a phase space reconstruction trajectory is constructed, and one component of the feature vector is selected as the basis time series for reconstruction, such as the valve opening component. The time delay coordinate method is used for phase space reconstruction, that is, the original time series and its delay sequence are combined into a multi-dimensional vector. Determining the appropriate time delay parameter and embedding dimension is the key to reconstruction.

[0116] The time delay parameter is determined, and the original time series is preprocessed, including detrending and normalization. The mutual information between the original sequence and the sequence with different delays is calculated. The curve of mutual information changing with delay is drawn. The first local minimum point of the curve is found, and the corresponding delay value is the appropriate time delay parameter. For example, for the pressure data of the first bypass valve, the calculated mutual information is 3.2 when the delay is 0, and gradually decreases with the increase of the delay, reaching the first local minimum value of 1.5 when the delay is 7. Therefore, the time delay parameter is selected as 7.

[0117] The embedding dimension is determined, and the initial embedding dimension d is set to 1. In the d-dimensional phase space, the nearest neighbor point of each point is found. The embedding dimension is increased to d+1, and it is checked whether the distance of the nearest neighbor point in the new dimension exceeds the threshold value. If it exceeds the threshold value, the point is marked as a false nearest neighbor point. The proportion of false nearest neighbor points is calculated. The embedding dimension is gradually increased, and the above process is repeated. When the proportion of false nearest neighbor points is lower than the preset threshold value (usually 5%), the corresponding embedding dimension is the appropriate embedding dimension. For example, for the flow data of the second bypass valve, when the embedding dimension is 1, 2, 3, 4, and 5, the proportions of false nearest neighbor points are 87%, 42%, 18%, 7%, and 3% respectively. Therefore, the embedding dimension is selected as 5.

[0118] According to the determined embedding dimension and time delay parameter, the phase space trajectory is reconstructed, for the original time series x(t), a d-dimensional vector [x(t), x(t+τ), x(t+2τ),..., x(t+(d-1)τ)] is constructed, where d is the embedding dimension and τ is the time delay parameter; for the original sequence with length N, N-(d-1)τ d-dimensional vectors can be obtained; the trajectory formed by these vectors in the d-dimensional space is the reconstructed phase space trajectory. For example, for a time series with length 10000, the time delay parameter τ=7 and the embedding dimension d=5 are selected, 9972 5-dimensional vectors can be obtained to form the phase space trajectory.

[0119] The Lyapunov exponent of the phase space trajectory is calculated, a reference point is selected in the phase space; a near neighbor point is found in its small neighborhood (usually with a radius of 1% of the diameter of the phase space); the evolution trajectories of the reference point and the near neighbor point are tracked, and the distance change between them with time is calculated; the natural logarithm of the distance change is taken, and a curve changing with time is drawn; linear fitting is performed on the curve, and the slope is the maximum Lyapunov exponent; repeat the above process to calculate all reference points, and take the average value as the maximum Lyapunov exponent of the system. For example, for the phase space trajectory of the first bypass valve, 1000 reference points are selected for calculation, and the maximum Lyapunov exponent obtained is 0.038, indicating that the system has weak chaotic characteristics.

[0120] Multi-scale fusion analysis is performed on the sliding entropy value and the Lyapunov exponent, and the scale factor range is set to 1-20; for each scale factor s, the original time series is divided into non-overlapping segments with length s; the average value of each segment is calculated to obtain a coarse-grained sequence; the sliding entropy value and the Lyapunov exponent of the coarse-grained sequence are calculated; a two-dimensional matrix is constructed, with rows representing different scales and columns representing different time points; the matrix elements are normalized to make their range between 0 and 1. For example, for the second bypass valve, when the scale factor s=1 (i.e. the original scale), the sliding entropy value is 0.76 and the Lyapunov exponent is 0.042; when s=5, the sliding entropy value decreases to 0.58 and the Lyapunov exponent decreases to 0.028; when s=10, the sliding entropy value further decreases to 0.45 and the Lyapunov exponent decreases to 0.021; this indicates that as the observation scale increases, the complexity and instability of the system decrease.

[0121] A cooperative evolution matrix of the dual bypass valve state characteristics is constructed, and the sliding entropy values and Lyapunov indexes at multiple scales are organized in the form of a matrix. The rows of the matrix represent different scale factors, and the columns represent different time points. The element values of the matrix are the weighted sums of the normalized entropy values and Lyapunov indexes, and the weights can be set according to expert experience, usually 50% each. For example, for 20 scale factors and 48 time points, a cooperative evolution matrix with a dimension of 20×48 is constructed. For the first bypass valve, when the scale factor s=3 and the time point t=24, the normalized sliding entropy value is 0.65, and the Lyapunov index is 0.72. Therefore, the corresponding element value of the cooperative evolution matrix is 0.65×0.5+0.72×0.5=0.685.

[0122] The cooperative evolution matrix is subjected to eigenvalue decomposition, and the covariance matrix of the cooperative evolution matrix is calculated. For a matrix with a dimension of 20×48, the covariance matrix has a dimension of 20×20. The covariance matrix is subjected to eigenvalue decomposition to obtain 20 eigenvalues and corresponding eigenvectors. The eigenvalues are arranged in descending order to obtain an eigenvalue sequence. The contribution rates and cumulative contribution rates of the eigenvalues are calculated. The first few eigenvalues with a cumulative contribution rate reaching a preset threshold (usually 90%) are selected as dominant eigenvalues. For example, for the cooperative evolution matrix of the second bypass valve, the first five eigenvalues obtained by eigenvalue decomposition are 4.82, 2.15, 0.93, 0.45, and 0.21, respectively, and their contribution rates are 56.2%, 25.1%, 10.8%, 5.3%, and 2.5%, respectively. The cumulative contribution rates of the first three eigenvalues reach 92.1%. Therefore, the first three eigenvalues are selected as dominant eigenvalues.

[0123] A state evolution feature space is constructed based on the dominant eigenvalues, and the eigenvectors corresponding to the dominant eigenvalues are extracted. These eigenvectors constitute the basis of the state evolution feature space. The cooperative evolution matrix is projected onto these eigenvectors to obtain projection coordinates. These projection coordinates constitute the trajectory in the state evolution feature space. For example, for the first bypass valve, the eigenvectors corresponding to the first three dominant eigenvalues are selected to construct a three-dimensional state evolution feature space. The cooperative evolution matrix is projected onto these three eigenvectors to obtain 48 three-dimensional coordinate points, which constitute the state trajectory.

[0124] The convergence radius of the state trajectory is calculated in the state evolution feature space, and the attractor of the state trajectory is identified. The highest density region in the state space can be identified as the attractor. The distances from the trajectory points to the attractor are calculated. The mean and standard deviation of the distances are calculated. The convergence radius is defined as the interval composed of the mean ± 3×standard deviation. For example, for the second bypass valve, the identified attractor coordinates are (0.15, 0.08, 0.03). The distances from the 48 trajectory points to the attractor are calculated, and the mean is 0.12 and the standard deviation is 0.03. Therefore, the convergence radius is [0.03, 0.21].

[0125] The divergence coefficient of the state trajectory is calculated, a pair of state vectors at adjacent time points is selected, the Euclidean distance between the vector pair is calculated, the rate of change of the distance with time, i.e. the ratio of the distances of the two adjacent points, is calculated, and the mean value of the rate of change is calculated, i.e. the divergence coefficient. For example, for the first bypass valve, the distance change rates of 47 adjacent vector pairs are calculated, and the mean value is 0.025, which is the divergence coefficient.

[0126] The state evolution trend characteristics of the industrial cooler double bypass valve are determined according to the convergence radius and the divergence coefficient, the state trend judgment standard is established, for example, the convergence radius is less than 0.2 and the divergence coefficient is less than 0.05, indicating a stable state, and the convergence radius is greater than 0.2 or the divergence coefficient is greater than 0.05, indicating an unstable state; the calculated convergence radius and divergence coefficient are compared with the judgment standard to determine the current state; the change trend of the convergence radius and the divergence coefficient is tracked to predict the future state evolution direction. For example, for the second bypass valve, if the convergence radius gradually changes from the initial [0.03, 0.21] to [0.08, 0.32], and the divergence coefficient gradually increases from 0.025 to 0.048, it indicates that the system state is transitioning from stable to unstable, i.e. a fault will occur.

[0127] In practical applications, the threshold standard for state trend judgment can be established based on historical data, for example, by analyzing a large number of normal operation and fault state data, the critical values of the convergence radius and the divergence coefficient are determined. For a certain type of industrial cooler double bypass valve, 500 sets of historical data are analyzed, including 400 sets of normal data and 100 sets of fault data, the statistical results show that: in normal state, the mean value of the upper limit of the convergence radius is 0.19, the standard deviation is 0.03, the mean value of the divergence coefficient is 0.028, and the standard deviation is 0.007; in fault state, the mean value of the upper limit of the convergence radius is 0.36, the standard deviation is 0.08, the mean value of the divergence coefficient is 0.072, and the standard deviation is 0.015. Based on these statistical results, the critical value of the upper limit of the convergence radius is set to 0.25, and the critical value of the divergence coefficient is set to 0.05, when the measured value exceeds these critical values, the system sends a warning signal.

[0128] The application scenarios of state evolution trend characteristics include fault warning and life prediction, for example, the first bypass valve of a certain industrial cooling system has been running continuously for 6 months, its state trajectory convergence radius gradually expands from the initial [0.05, 0.17] to [0.09, 0.28], and the divergence coefficient increases from 0.022 to 0.058, exceeding the preset critical value, the system automatically generates a warning message. The maintenance personnel check the valve and find that the sealing surface has a small wear and a slight leakage, so the sealing element is replaced during the planned shutdown, avoiding sudden shutdown due to valve failure, saving maintenance cost and improving system reliability.

[0129] In an optional embodiment, the sliding entropy value and the Lyapunov exponent are subjected to multi-scale fusion analysis to construct a cooperative evolution matrix of the state characteristics of the dual bypass valve.

[0130] A bivariate cooperative matrix is constructed according to the sliding entropy value and the Lyapunov exponent, and recursive quantization analysis is performed on the bivariate cooperative matrix to establish a dynamic association network of state characteristics.

[0131] The topological features and transmission characteristics of the dynamic association network are extracted, and the coupling relationship between state characteristics is determined according to the topological features and the transmission characteristics.

[0132] The coupling relationship is subjected to wavelet transform to obtain wavelet coefficients at different scales, and the wavelet coefficients are divided into high-frequency subbands and low-frequency subbands according to the energy distribution law.

[0133] A multi-scale feature matrix is constructed based on the high-frequency subbands and the low-frequency subbands, and principal component feature vectors are obtained by singular value decomposition of the multi-scale feature matrix.

[0134] The principal component feature vectors are projected into a preset state evolution space to obtain a state evolution projection matrix, and a cooperative evolution matrix of the state characteristics of the dual bypass valve is constructed according to the state evolution projection matrix.

[0135] The sliding entropy value and the Lyapunov exponent are subjected to multi-scale fusion analysis to construct a cooperative evolution matrix of the state characteristics of the dual bypass valve. The sliding entropy value reflects the complexity of the system, and the Lyapunov exponent represents the degree of chaos of the system. The combination of the two can comprehensively describe the dynamic characteristics of the system. Multi-scale fusion analysis considers the behavior differences of the system at different time scales and spatial scales, which helps to reveal the internal laws of the state evolution of the system.

[0136] A bivariate cooperative matrix is constructed according to the sliding entropy value and the Lyapunov exponent, and recursive quantization analysis is performed on the bivariate cooperative matrix to establish a dynamic association network of state characteristics. The sliding entropy value sequence is denoted as H(t), the Lyapunov exponent sequence is denoted as L(t), t represents the time point, and the range is 1 to T; the covariance between H(t) and L(t) at different time points is calculated to form a T×T covariance matrix; the covariance matrix is normalized to obtain a correlation coefficient matrix, i.e. a bivariate cooperative matrix. For example, for a first bypass valve of an industrial cooling system, 72 hours of monitoring data are collected, with a time resolution of 1 hour, and a 72×72 bivariate cooperative matrix is obtained. The matrix element value ranges from -1 to 1, a positive value indicates a positive correlation, a negative value indicates a negative correlation, and the absolute value size indicates the correlation strength.

[0137] Set the recursion depth D (usually 3-5) and the threshold parameter a (usually 0.7); for each element of the correlation matrix, if its absolute value is greater than a, a connection is established between the corresponding time points to form an initial network (depth 1); for a network with depth d, check whether there is a path of length d between any two nodes, if there is, a new connection is established between the two nodes to form a network with depth d+1; repeat the above process until the set recursion depth D is reached. For example, for the above correlation matrix, set a=0.7, D=3, through recursive quantification analysis, a dynamic association network containing 72 nodes is obtained, where the nodes represent different time points and the edges represent the association relationship between time points.

[0138] Extract the topological features and transfer characteristics of the dynamic association network, and determine the coupling relationship between the state features according to the topological features and the transfer characteristics. Topological features mainly include node degree distribution, clustering coefficient, average path length, etc., and transfer characteristics mainly include information transfer efficiency, robustness, etc. Calculate the degree of each node in the network (the number of edges directly connected to it) to obtain the degree distribution; calculate the global clustering coefficient of the network to reflect the aggregation degree of the network; calculate the average path length of the network to reflect the transmission efficiency of the network; analyze the robustness of the network by randomly removing nodes. For example, for the dynamic association network of the second bypass valve, the calculated average node degree is 12.5, the global clustering coefficient is 0.68, and the average path length is 2.8, indicating that the network has high aggregation and efficiency.

[0139] Identify the community structure in the network through community detection algorithm, the nodes in the same community have strong coupling relationship; calculate the structural similarity between nodes, the nodes with high similarity correspond to state features with similar coupling mode; analyze the hierarchical structure of the network to determine the core nodes and edge nodes, the state features corresponding to the core nodes play a leading role in system evolution. For example, for the dynamic association network of the first bypass valve, three main communities are identified through community detection, corresponding to the startup stage, stable running stage and load change stage; the core nodes are mainly distributed in the stable running stage, indicating that the state features of this stage have a great influence on the overall behavior of the system.

[0140] Wavelet transform is performed on the coupling relationship to obtain wavelet coefficients at different scales, and the wavelet coefficients are divided into a high-frequency subband and a low-frequency subband according to an energy distribution rule. The coupling relationship can be represented as a time series to describe the change of the correlation strength between state characteristics over time. A suitable wavelet basis function, such as db4 or sym8, is selected, and the decomposition layer number L is set (usually 3-5). Wavelet decomposition is performed on the time series of the coupling relationship to obtain L layers of detail coefficients and 1 layer of approximation coefficients. The energy distribution of each layer of coefficients is calculated. For example, for a certain time series of the coupling relationship, 4 layers of decomposition are performed using the db4 wavelet, and the energy distribution obtained is as follows: the first layer (highest frequency) accounts for 10%, the second layer accounts for 15%, the third layer accounts for 25%, the fourth layer accounts for 20%, and the approximation coefficient (lowest frequency) accounts for 30%.

[0141] The cumulative energy distribution is calculated, an energy threshold β is set (usually 0.5-0.7), and the energy is accumulated from high frequency to low frequency. When the cumulative energy reaches β, the frequency band is used as the demarcation line between the high-frequency subband and the low-frequency subband. For the above example, assuming β=0.5, the cumulative energy of the first layer and the second layer is 25%, and the energy of the third layer is 25%, which accumulates to 50%. Therefore, the first layer to the third layer are divided into a high-frequency subband (accounting for 50% of the total energy), and the fourth layer and the approximation coefficient are divided into a low-frequency subband (accounting for 50% of the total energy). The high-frequency subband mainly reflects the rapid change and transient characteristics of the system, and the low-frequency subband mainly reflects the slow change and long-term trend of the system.

[0142] Based on the high-frequency subband and the low-frequency subband, a multi-scale feature matrix is constructed, and principal component feature vectors are obtained by singular value decomposition of the multi-scale feature matrix. The wavelet coefficients of the high-frequency subband and the low-frequency subband are reconstructed into time series, and the reconstructed sequences are processed in segments, with each segment having a length of W (usually 10%-20% of the original sequence length). Statistical features such as mean, standard deviation, and kurtosis are calculated for each segment. These statistical features are organized into a matrix form, with rows representing different time periods and columns representing different statistical features. For example, for 72 hours of monitoring data, setting W=12 hours can obtain 6 time periods, and 10 statistical features are calculated for each time period to form a 6x10 multi-scale feature matrix.

[0143] The multi-scale feature matrix is denoted as M, and singular value decomposition is performed on M to obtain three matrices U, S, and V, where S is a diagonal matrix and the elements on the diagonal are singular values. The left singular vectors corresponding to the first K largest singular values are selected according to the size of the singular values to form a principal component feature vector matrix P. For example, for a 6x10 multi-scale feature matrix, singular value decomposition obtains 6 singular values, which are 4.25, 2.18, 1.03, 0.56, 0.32, and 0.11, respectively. The cumulative contribution rate of the first three singular values is 88.7%, so the first three left singular vectors are selected to form a 3x6 principal component feature vector matrix.

[0144] The principal component feature vector is projected to a preset state evolution space to obtain a state evolution projection matrix, and a cooperative evolution matrix of the state features of the dual bypass valve is constructed according to the state evolution projection matrix. The preset state evolution space is a standard space determined based on expert knowledge and historical data analysis, and is usually a low-dimensional space (2-4 dimensions). A projection matrix Q is designed, with a dimension of KxR, where K is the dimension of the principal component feature vector, and R is the dimension of the preset state evolution space; a state evolution projection matrix E with a dimension of Rx6 is obtained by matrix multiplication PxQ; and the E is normalized to make the value range of each element between 0 and 1. For example, for a principal component feature vector matrix of 3x6, a projection matrix of 3x2 is designed, a state evolution projection matrix of 2x6 is obtained, and the projection coordinates of the six time periods in the two-dimensional state evolution space are represented.

[0145] The state distance between different time periods is calculated to form a distance matrix; the distance matrix is converted into a similarity matrix; and the similarity matrix is normalized to obtain a cooperative evolution matrix. For example, for a state evolution projection matrix of 2x6, a distance matrix of 6x6 is calculated, and the element values represent the Euclidean distances between the states of different time periods; the distance is converted into similarity (similarity = 1 / (1+distance)), and a similarity matrix of 6x6 is obtained; and the similarity matrix is row-normalized to obtain the final cooperative evolution matrix. The element values of the cooperative evolution matrix represent the transition probabilities between the states of different time periods, reflecting the internal law of system state evolution.

[0146] In practical applications, the construction of the bivariate cooperative matrix needs to consider the stationarity and correlation of the time series. For non-stationary time series, difference or detrending processing is needed. When calculating the covariance, the sliding window technique can be used to reduce the influence of local fluctuations. For the sliding entropy value sequence and the Lyapunov index sequence of the first bypass valve, the sequence is first standardized (subtract the mean value and divide by the standard deviation) before calculating the covariance; then a 30-point sliding window is used to calculate the local covariance; and finally the local covariance is combined into a global covariance matrix. For 72 hours of data, the standardized sliding entropy value sequence has a mean of 0 and a standard deviation of 1, with a value range of -2.8 to 3.2; the Lyapunov index sequence has a mean of 0 and a standard deviation of 1, with a value range of -2.5 to 2.9.

[0147] The selection of threshold parameter a has a significant impact on the network structure in the recursive quantization analysis process. A smaller a value will lead to an overly dense network connection, which cannot reflect the true association structure. A larger a value will lead to an overly sparse network, losing important information. By comparing the network topologies generated by different a values, the optimal threshold can be determined. For the second bypass valve's bivariate coordination matrix, when a = 0.5, the generated network has an average node degree of 35.2 and a clustering coefficient of 0.92, indicating an overly dense network structure. When a = 0.9, the average node degree is 3.8 and the clustering coefficient is 0.42, indicating an overly sparse network structure. When a = 0.7, the average node degree is 12.5 and the clustering coefficient is 0.68, indicating a moderate network structure that can reasonably reflect the system's association characteristics.

[0148] The extraction of topological features of dynamic association networks needs to consider the size and complexity of the network. For large-scale networks, sampling techniques can be used to reduce computational complexity. The transfer property analysis of the network can be achieved by simulating the information flow propagation process. For the dynamic association network of the first bypass valve, the average first arrival time (the average number of steps required to reach one node from another) between nodes is calculated by simulating the random walk process, and the transfer efficiency matrix is constructed. For a network of 72 nodes, the average first arrival time is 4.2 steps, and the maximum first arrival time is 12 steps, indicating that the network has high transfer efficiency.

[0149] In the wavelet transform process, the selection of wavelet basis function should consider the characteristics of the signal. For a relatively smooth signal, it is suitable to select a wavelet basis with good symmetry, such as sym series. For signals containing mutations, it is suitable to select a wavelet basis with good orthogonality, such as db series. The determination of the decomposition level can be based on the signal length and frequency characteristics, usually following the principle of L ≤ log2(N), where N is the signal length. For a coupling relationship time series with a length of 1024, the maximum decomposition level is 10. Considering the balance between computational efficiency and information preservation, the decomposition level is selected as 4. Using db4 wavelet for decomposition, the first layer frequency band is 256-512Hz, the second layer is 128-256Hz, the third layer is 64-128Hz, and the fourth layer is 32-64Hz. The approximation coefficients correspond to 0-32Hz.

[0150] In the construction of the multiscale feature matrix, the choice of the time period length W needs to balance the time resolution and statistical reliability. A smaller W value provides higher time resolution but leads to unstable statistical features; a larger W value provides more reliable statistical features but reduces the time resolution. The choice of statistical features should cover different aspects of the signal, including central tendency (mean, median), dispersion (standard deviation, range), distribution shape (kurtosis, skewness), etc. For the reconstructed sequences of the high-frequency subband and the low-frequency subband of the second bypass valve, 10 statistical features are calculated: mean, standard deviation, maximum, minimum, range, median, interquartile range, kurtosis, skewness, and zero-crossing rate. Among them, the kurtosis value of the high-frequency subband is generally higher than that of the low-frequency subband, indicating that the high-frequency component contains more transient mutation information.

[0151] In singular value decomposition, the determination of the dimension K of the principal component feature vector is usually based on the cumulative contribution rate, and the first K singular values with a cumulative contribution rate of 85%-95% are selected. For different types of valves and working conditions, the optimal K value is different. For the first bypass valve in a stable running state, the cumulative contribution rate of the first 2 singular values is 92.3%; for the second bypass valve with frequent load changes, the first 4 singular values are needed to reach the same contribution rate, indicating that the state change of the latter is more complex.

[0152] The determination of the dimension R of the state evolution space needs to consider the visualization needs and information retention needs. For visualization purposes, R = 2 or R = 3 is usually chosen; for the need to retain more information, a higher dimension can be chosen. The design of the projection matrix Q can be based on principal component analysis or manifold learning algorithms, with the goal of preserving the structure information of the original data to the maximum extent while reducing the dimension. For the first bypass valve, a 3x2 projection matrix is designed to project the 3-dimensional principal component feature vector into a 2-dimensional state evolution space. The projected points on the plane show a spiral-shaped trajectory, clearly showing the periodic changes and long-term evolution trend of the system state.

[0153] The construction of the co-evolution matrix is the final step of multiscale fusion analysis, which describes the transition relationship between different time periods of the system state. By analyzing the characteristics of the co-evolution matrix, the future state of the system can be predicted and abnormal behavior can be identified. For the normally operating double bypass valve, its co-evolution matrix has good symmetry and strong diagonality, indicating that the system state is relatively stable; when the valve shows signs of failure, the symmetry of the co-evolution matrix decreases and the non-diagonal element value increases, indicating that the system state becomes unstable. After 72 hours of operation, the symmetry index of the co-evolution matrix of the second bypass valve of a certain industrial cooling system decreased from 0.92 to 0.76, and the system timely issued a warning, and maintenance personnel found that the actuator of the valve was loose and timely adjusted, avoiding more serious failure.

[0154] In an alternative embodiment, a fault feature index of the industrial cooler double bypass valve is calculated according to the state evolution trend feature and the deviation value of the valve opening degree in the operating state feature vector from the input control signal, comprising:

[0155] A dynamic transfer matrix is constructed based on the state evolution trend feature, and a valve response feature matrix is constructed based on the deviation value of the valve opening degree in the operating state feature vector from the input control signal;

[0156] The dynamic transfer matrix and the valve response feature matrix are cooperatively mapped to obtain a dynamic-static feature transfer path of the double bypass valve and a valve response topological correlation degree;

[0157] The valve opening degree distribution characteristics of the double bypass valve under different operating states are extracted according to the dynamic-static feature transfer path, and a spatial mapping matrix of the valve opening degree is constructed;

[0158] The opening degree response correlation strength and the opening degree synchronization index of the double bypass valve are calculated based on the valve response topological correlation degree, and a correlation mapping matrix of the valve opening degree is constructed;

[0159] The spatial mapping matrix of the valve opening degree and the correlation mapping matrix of the valve opening degree are feature fused to obtain a valve opening degree comprehensive evaluation index of the double bypass valve, and a fault feature index of the industrial cooler double bypass valve is calculated according to the valve opening degree comprehensive evaluation index.

[0160] A fault feature index of the industrial cooler double bypass valve is calculated according to the state evolution trend feature and the deviation value of the valve opening degree in the operating state feature vector from the input control signal. The state evolution trend feature reflects the dynamic change characteristics of the valve state, and the deviation value of the valve opening degree from the input control signal directly reflects the accuracy of the valve response. The combined analysis of these two types of information can comprehensively evaluate the health state of the valve.

[0161] Based on the state evolution trend characteristics, a dynamic transfer matrix is constructed, and a valve opening degree and an input control signal deviation value in the operating state feature vector construct a valve response feature matrix. The dynamic transfer matrix describes the transfer relationship between system states at different time points, and the valve response feature matrix describes the response characteristics of the valve to the control signal. The process of constructing the dynamic transfer matrix is as follows: extracting the convergence radius and divergence coefficient in the state evolution trend characteristics; organizing the values of the convergence radius and the divergence coefficient at different time points into vector form; calculating the transfer probability between the vectors to form a probability transfer matrix; and normalizing the probability transfer matrix to obtain the dynamic transfer matrix. For example, for a first bypass valve of an industrial cooling system, state data is collected once a day for 10 days, and 10 sets of convergence radius and divergence coefficient values are obtained. The transfer relationship between adjacent time points is calculated, and a 10*10 dynamic transfer matrix is constructed, and the matrix element value represents the probability of state transfer from one time point to another time point.

[0162] The valve response feature matrix is constructed, the deviation value of the valve opening degree and the input control signal is collected, and is recorded as a deviation sequence; the deviation sequence is processed in segments, and each segment has a length L (usually 24 hours); statistical characteristics in each segment are calculated, such as mean, standard deviation, maximum value, minimum value, etc.; and the statistical characteristics are organized into a matrix form, with rows representing different time periods and columns representing different statistical characteristics. For example, for the second bypass valve, 10 days of deviation data are collected, each day is taken as a time period, and 6 statistical characteristics are calculated to obtain a 10*6 valve response feature matrix. In the normal operating state, the deviation mean is close to 0, and the standard deviation is small (such as 0.5%); when the valve fails, the deviation mean deviates significantly from 0 (such as 2.5%), and the standard deviation increases (such as 1.8%).

[0163] The dynamic transfer matrix and the valve response feature matrix are cooperatively mapped to obtain the dynamic-static feature transfer path and the valve response topological correlation degree of the double bypass valve. The purpose of cooperative mapping is to fuse the state evolution information and the valve response information, and to reveal the internal relationship between the two. The specific steps of cooperative mapping include: standardizing the dynamic transfer matrix and the valve response feature matrix; calculating the mutual information matrix of the two matrices; constructing a bipartite graph based on the mutual information matrix, with nodes representing state features and response features respectively, and the weight of the edge representing the mutual information amount between the features; performing community detection on the bipartite graph to identify closely related feature groups; extracting the connection path in the community to form the dynamic-static feature transfer path; and calculating the connection density between the communities to obtain the valve response topological correlation degree.

[0164] For example, for a certain dual bypass valve system, the bipartite graph contains 10 state nodes and 6 response nodes, and 3 feature groups are identified by community detection, corresponding to the startup phase, stable operation phase and load change phase respectively. Within each group, the connections between state features and response features form feature transfer paths, such as "convergence radius - deviation standard deviation - divergence coefficient". The connection density between groups is 0.35, indicating a moderate degree of topological association.

[0165] According to the dynamic and static feature transfer paths, the valve opening distribution characteristics of the dual bypass valve under different operating conditions are extracted, and a spatial mapping matrix of valve opening is constructed. The valve opening distribution characteristics reflect the working state distribution of the valve under different working conditions. The steps of extracting the valve opening distribution characteristics include: collecting valve opening data along the feature transfer path; performing probability density estimation on the valve opening data to obtain a probability distribution function; calculating feature parameters of the distribution function, such as mean, variance, skewness, kurtosis, etc.; organizing these feature parameters into a matrix form, with rows representing different operating conditions and columns representing different feature parameters, to form a spatial mapping matrix. For example, for the first bypass valve, the mean of the valve opening distribution under the startup, stable operation and load change conditions is 25%, 60% and 40% respectively, and the standard deviation is 8%, 3% and 12% respectively, forming a 3x2 spatial mapping matrix. The opening distribution of a normal valve usually presents a single peak characteristic, with a kurtosis close to 3; while the opening distribution of a faulty valve presents a double peak or multi-peak characteristic, with a kurtosis significantly deviating from 3.

[0166] Based on the valve response topological association degree, the opening response association strength and the opening synchronization index of the dual bypass valve are calculated, and a correlation mapping matrix of valve opening is constructed. The opening response association strength represents the degree of association between the valve opening and the control signal, and the opening synchronization index represents the coordination degree of the opening change of the first bypass valve and the second bypass valve. The steps of calculating the opening response association strength include: extracting the time series of the control signal and the valve opening; calculating the cross-correlation function between the two sequences; finding the maximum value of the cross-correlation function and its corresponding time delay; the maximum value represents the association strength, and the time delay represents the response lag. The steps of calculating the opening synchronization index include: extracting the opening time series of the first bypass valve and the second bypass valve; calculating the cross-correlation function between the two sequences; calculating the correlation coefficient at zero delay point, which is the synchronization index. The association strength and the synchronization index are organized into a matrix form to form a correlation mapping matrix. For example, for a certain dual bypass valve system, the opening response association strength under three operating conditions is 0.92, 0.95 and 0.88 respectively, and the opening synchronization index is 0.85, 0.93 and 0.78 respectively, forming a 3x2 correlation mapping matrix.

[0167] The spatial mapping matrix of the valve opening degree and the correlation mapping matrix of the valve opening degree are feature fused to obtain a comprehensive evaluation index of the valve opening degree of the double bypass valve, and a fault feature index of the double bypass valve of the industrial cooler is calculated according to the comprehensive evaluation index of the valve opening degree. The purpose of feature fusion is to comprehensively consider the spatial distribution characteristics and the correlation characteristics to obtain a more comprehensive evaluation index. The steps of feature fusion include: standardizing the spatial mapping matrix and the correlation mapping matrix; setting a weight coefficient, usually 50% for the spatial characteristics and the correlation characteristics; calculating the weighted sum to obtain the comprehensive evaluation index; and normalizing the comprehensive evaluation index to make its range between 0 and 1.

[0168] A reference value of the normal state is set, a deviation between the comprehensive evaluation index of the current state and the reference value is calculated, the deviation is mapped to a range of 0-100 to obtain the fault feature index, wherein 0 represents complete health and 100 represents complete failure. For example, the comprehensive evaluation index of a certain double bypass valve system in the normal state is 0.92, the comprehensive evaluation index of the current state is 0.75, the deviation is 0.17, and the corresponding fault feature index is 35, indicating that the system has a slight fault.

[0169] In the process of constructing the dynamic transfer matrix, the extraction of the convergence radius and the divergence coefficient needs to consider the stability and representativeness of the data. For each monitoring period, the convergence radius and the divergence coefficient can be calculated multiple times to improve the stability by taking their average value. For example, for the first bypass valve, 24 groups of hourly data are collected within a day, the convergence radius and the divergence coefficient are calculated respectively, and then the average value is taken as the representative value of the day. When calculating the transfer probability, the similarity of the states needs to be considered, and the similarity between the state vectors can be calculated by using a Gaussian kernel function. For example, when the Euclidean distance between two state vectors is 0.05, the corresponding similarity is 0.95; when the distance is 0.2, the similarity is 0.82; and when the distance is 0.5, the similarity is 0.61. Based on these similarity values, a transition probability matrix between states can be constructed.

[0170] In the process of constructing the valve response feature matrix, the calculation of the deviation value needs to consider the time alignment of the control signal and the valve opening. Due to the certain hysteresis of the valve response, directly calculating the difference between the control signal and the opening value at the same time is not accurate. A solution is to introduce a time window to find the best match between the opening value and the control signal within the window. For example, for the second bypass valve with a response time of about 2 seconds, a 5-second time window is set to find the minimum deviation between the opening value and the control signal within the window, which is taken as the actual response deviation. The selection of statistical characteristics should cover different aspects of the deviation, including central tendency, dispersion, and distribution shape. For example, the mean deviation of a certain fault valve is 2.5%, the standard deviation is 1.8%, the maximum deviation is 5.7%, the minimum deviation is 0.3%, the skewness is 0.8, and the kurtosis is 4.2, which together reflect the abnormality of the valve response.

[0171] In the process of collaborative mapping, the calculation of the mutual information matrix needs to consider the nonlinear relationship between features. Mutual information can capture the nonlinear dependence between variables, and is more universal than correlation coefficients. When calculating mutual information, the joint probability distribution and the marginal probability distribution need to be estimated, and histogram method or kernel density estimation method can be used. For example, for the two features of convergence radius and deviation standard deviation, their value ranges are divided into 10 intervals, the number of samples falling into each interval combination is counted, and a 10x10 joint frequency matrix is obtained. The mutual information value calculated based on this matrix is 0.65, indicating that there is a strong dependence between the two features. The selection of community detection algorithm should consider the characteristics of bipartite graph, and common algorithms include label propagation algorithm and modularity optimization algorithm. For example, for a bipartite graph containing 16 nodes, 3 communities are detected by the label propagation algorithm, and the modularity is 0.58, indicating that the community structure is significant.

[0172] In the process of extracting the valve opening distribution characteristics, the method selection of probability density estimation should consider the distribution characteristics of the data. For data with approximately normal distribution, parameter estimation method can be used; for data with complex distribution, kernel density estimation method or histogram method can be used. For example, for the opening data of the second bypass valve in the stable running state, the probability density function estimated by the Gaussian kernel density estimation presents a single peak characteristic, the peak value is located at the opening of 65%, and the 90% confidence interval of the distribution is [58%, 72%], indicating that the valve opening is stable. For the fault valve, the opening distribution presents a double peak characteristic, one peak is near the target opening, and the other peak is at a lower or higher opening value, reflecting the existence of valve sticking or drifting phenomenon.

[0173] In the process of calculating the correlation strength of the opening response, the length of the calculation window of the cross-correlation function should be determined according to the dynamic characteristics of the system. For valves with faster response, a shorter window can be used; for valves with slower response, a longer window should be used. For example, for the first bypass valve with a response time of about 1 second, a 10-second calculation window is used; for the second bypass valve with a response time of about 3 seconds, a 30-second calculation window is used. The cross-correlation function of a normal valve usually presents a clear peak, and the peak position corresponds to the response delay, and the peak size is close to 1; the cross-correlation function of a fault valve has a lower peak value, or multiple peaks appear, indicating unstable response. For example, for a normally operating first bypass valve, the maximum value of its cross-correlation function is 0.95, and the corresponding time delay is 1.2 seconds; while the maximum value of the cross-correlation function of the fault valve is only 0.72, and the corresponding time delay is 2.8 seconds, indicating that the response is lagging and inaccurate.

[0174] In the calculation of the opening synchronization index, the coordinated working characteristics of the first bypass valve and the second bypass valve need to be considered. In some working conditions, the two valves need to change in the same direction, at which time the synchronization index should be close to 1; in other working conditions, the two valves need to change in opposite directions, at which time the synchronization index should be close to -1. For example, during the system startup phase, the two valves need to be opened at the same time, and the normal synchronization index is 0.92; during the load balancing phase, the two valves need to be replaced by each other, and the normal synchronization index is -0.85. In the fault condition, the synchronization index deviates significantly from the expected value, such as in the working condition that should change in the same direction, the synchronization index drops to 0.35, indicating poor valve coordination.

[0175] In the feature fusion process, the determination of the weight coefficient can be based on expert experience or historical data analysis, and the relative importance of spatial characteristics and correlation characteristics varies with system configuration and operating conditions. For example, for a system with large flow fluctuations, the correlation characteristics are more important, and the weight can be set to 60%; for a stable running system, the spatial characteristics are more important, and the weight can be set to 60%. The calculation of the comprehensive evaluation index should consider the correlation between the features to avoid the influence of repeated calculation of similar features. For example, for a double bypass valve system, the score of the spatial mapping matrix is 0.88, and the score of the correlation mapping matrix is 0.81, considering the correlation coefficient between the two is 0.3, the adjusted comprehensive evaluation index is 0.85.

[0176] The calculation of the fault feature index requires establishing a reasonable mapping relationship so that the index value can intuitively reflect the severity of the fault. A piecewise function can be used, which gives a lower fault index for slight deviations and a higher fault index for severe deviations. For example, when the deviation of the comprehensive evaluation index is between 0 and 0.1, the fault feature index linearly varies between 0 and 20; when the deviation is between 0.1 and 0.3, the fault feature index linearly varies between 20 and 60; and when the deviation is greater than 0.3, the fault feature index linearly varies between 60 and 100. This segmented mapping can highlight the key deviation range and improve the sensitivity of fault diagnosis. In actual application, the fault feature index of a double bypass valve in an industrial cooling system gradually rises from 5 to 45 after 3 months of operation, and the system issues a warning signal. Maintenance personnel find that the sealing surface of the second bypass valve is worn out, and the sealing element is replaced in time to avoid more serious faults.

[0177] In this embodiment, the fault feature index not only reflects the current fault state, but also predicts potential fault risks through trend analysis. By continuously monitoring the change trend of the fault feature index, the development speed and severity of the fault can be identified. For example, the fault feature index of a double bypass valve system slowly rises from 15 to 25 within a week, indicating a slight fault but slow development; while the fault feature index of another system rapidly rises from 20 to 50 within two days, indicating that the fault is rapidly deteriorating and needs immediate intervention. This trend analysis capability provides important support for predictive maintenance, helping maintenance personnel take measures before the fault causes serious impact.

[0178] In an alternative embodiment, based on the historical data distribution variance of the fault feature index, combined with the operating condition parameters of the industrial cooler, the fault warning threshold of the fault feature index is determined, comprising:

[0179] Performing distribution characteristic analysis on the fault feature index to construct a historical distribution feature vector of the fault feature index;

[0180] Calculating the distribution skewness and distribution kurtosis of the fault feature index according to the historical distribution feature vector to obtain the historical data distribution variance of the fault feature index;

[0181] Classifying the operating condition parameters according to the condition type to construct a condition feature vector, and calculating the correlation coefficient and contribution coefficient of each parameter in the condition feature vector;

[0182] Constructing a condition parameter combination coefficient matrix based on the correlation coefficient and the contribution coefficient, and constructing a parameter distribution matrix based on the historical data distribution variance;

[0183] The parameter distribution matrix is coupled and mapped with the working condition parameter combination coefficient matrix, the correlation characteristics of the fault feature index and the working condition parameters are extracted, an adaptive threshold mapping function is constructed based on the correlation characteristics, and a fault warning threshold of the fault feature index is determined according to the adaptive threshold mapping function.

[0184] The fault warning threshold of the fault feature index is determined based on the historical data distribution variance of the fault feature index and in combination with the running working condition parameters of the industrial cooler. The fault feature index is a comprehensive index reflecting the health status of the double bypass valve, the historical data distribution characteristics of which contain the rules of normal fluctuations and abnormal changes of the system, and the running working condition parameters reflect the working state of the system, and the combination of the two can realize adaptive fault warning for different working conditions.

[0185] The distribution characteristics of the fault feature index are analyzed, and a historical distribution feature vector of the fault feature index is constructed. The distribution characteristic analysis mainly investigates the statistical distribution characteristics of the fault feature index under normal operating conditions. The specific steps include collecting historical data of the fault feature index for a long enough period of time, usually 3-6 months of operating data; pre-processing the data, including abnormal value detection and removal, data normalization, etc.; dividing the data into multiple intervals, counting the frequency of data points in each interval, and constructing a frequency distribution histogram; calculating the statistical characteristics of the distribution, including mean, standard deviation, quantile, etc.; and combining these statistical characteristics into a historical distribution feature vector.

[0186] For example, for a double bypass valve of an industrial cooling system, 180 days of fault feature index data are collected, totaling 10800 data points (sampling interval is 24 minutes). After data preprocessing, the calculated feature vector includes: mean 15.3, standard deviation 3.8, minimum value 5.2, 25% quantile 12.6, median 15.1, 75% quantile 18.2, and maximum value 26.9. These statistical characteristics collectively describe the distribution characteristics of the fault feature index.

[0187] The distribution skewness and kurtosis of the fault feature index are calculated according to the historical distribution feature vector, and the historical data distribution variance of the fault feature index is obtained. Skewness measures the asymmetry of the distribution, and kurtosis measures the sharpness of the distribution, and the combination of the two can more comprehensively describe the shape characteristics of the distribution. The steps for calculating skewness include: calculating the third central moment based on historical data; dividing the third central moment by the cube of the standard deviation to obtain the skewness value.

[0188] The fourth-order central moment is calculated based on the historical data; the fourth-order central moment is divided by the fourth power of the standard deviation to obtain the kurtosis value; the kurtosis value is usually subtracted by 3, so that the kurtosis of the normal distribution is 0. The distribution variance of the historical data is the square of the standard deviation, which is directly extracted from the historical distribution feature vector. For example, for the fault feature index of the double bypass valve described above, the skewness is 0.18, indicating that the distribution is slightly right-skewed; the kurtosis is -0.25, indicating that the distribution is slightly flatter than the normal distribution; the distribution variance of the historical data is 14.44 (i.e., the square of the standard deviation 3.8). These indicators collectively reflect the fluctuation characteristics of the fault feature index.

[0189] The operating condition parameters are classified according to the operating condition types to construct an operating condition feature vector, and the correlation coefficient and the contribution degree coefficient of each parameter in the operating condition feature vector are calculated. Operating condition parameters usually include flow, pressure, temperature, load and other parameters, and different parameters have different effects on the fault feature index. The steps of classifying operating condition parameters by type include: identifying the main operating condition parameters of the system; classifying according to the physical meaning and functional characteristics of the parameters, such as fluid parameter class, thermal parameter class, mechanical parameter class, etc.; selecting representative parameters in each category to form an operating condition feature vector. The steps of calculating the correlation coefficient include: calculating the Pearson correlation coefficient of each operating condition parameter and the historical data of the fault feature index; the correlation coefficient takes a value in the range [-1, 1], and the greater the absolute value, the stronger the correlation.

[0190] The contribution degree coefficient is calculated to construct a regression model of the fault feature index and the operating condition parameters, and the regression coefficients of each parameter in the model are extracted; the regression coefficients are standardized to obtain the contribution degree coefficients. For example, the operating condition parameters of a certain industrial cooling system are divided into three categories: fluid parameters (including flow, pressure, flow rate), thermal parameters (including inlet and outlet temperatures, temperature difference), and mechanical parameters (including vibration, noise). Selecting flow, pressure, inlet and outlet temperature, temperature difference, and vibration intensity to form an operating condition feature vector. The correlation coefficients of these parameters and the fault feature index are calculated to be 0.65, 0.58, 0.42, 0.73, and 0.38, respectively, and the contribution degree coefficients are 0.25, 0.18, 0.12, 0.32, and 0.13, respectively.

[0191] Based on the correlation coefficient and the contribution coefficient, a working condition parameter combination coefficient matrix is constructed, and a parameter distribution matrix is constructed based on the historical data distribution variance. The working condition parameter combination coefficient matrix describes the combination relationship between different working condition parameters and the comprehensive influence on the fault characteristic index. The steps of constructing the working condition parameter combination coefficient matrix include: calculating the interaction coefficient between working condition parameters, that is, the influence of the common change of two parameters on the fault characteristic index; combining the correlation coefficient and the contribution coefficient to calculate the weight coefficient of parameter combination; and organizing the weight coefficient into a matrix form, with the rows and columns representing different working condition parameters. The parameter distribution matrix describes the distribution characteristics of the fault characteristic index under different working condition parameter values.

[0192] The parameter distribution matrix divides the value range of each working condition parameter into multiple intervals. For each parameter interval combination, the conditional distribution variance of the fault characteristic index under the corresponding condition is calculated, and the conditional distribution variance is organized into a matrix form. For example, for the two parameters of flow and temperature difference that have the greatest influence, the flow is divided into low, medium and high intervals, and the temperature difference is divided into small, medium and large intervals, to obtain a 3x3 parameter distribution matrix. Under the conditions of low flow and small temperature difference, the conditional distribution variance of the fault characteristic index is 10.2; under the conditions of high flow and large temperature difference, the conditional distribution variance is 18.7.

[0193] The parameter distribution matrix and the working condition parameter combination coefficient matrix are coupled and mapped to extract the correlation characteristics of the fault characteristic index and the working condition parameters, and an adaptive threshold mapping function is constructed based on the correlation characteristics. The purpose of the coupling and mapping is to establish a quantitative relationship between the working condition parameters and the distribution characteristics of the fault characteristic index. The steps of the coupling and mapping include: standardizing the parameter distribution matrix and the working condition parameter combination coefficient matrix; calculating the product of the two matrices to obtain a correlation characteristics matrix; and extracting the eigenvectors of the correlation characteristics matrix as the correlation characteristics. The adaptive threshold mapping function establishes a functional relationship between the working condition parameters and the fault warning threshold.

[0194] An adaptive threshold mapping function is constructed, a mapping relationship between the working condition parameters and the threshold adjustment factor is constructed based on the correlation characteristics, a reference threshold is set, which is usually the mean value of the fault characteristic index in the normal state plus multiple standard deviations, and the threshold adjustment factor is applied to the reference threshold to obtain an adaptive threshold for specific working conditions. For example, the reference threshold of a certain double bypass valve is set to the mean value 15.3 plus 3 times the standard deviation 3.8, that is, 26.7. Through the adaptive threshold mapping function, in the working condition of low flow and small temperature difference, the threshold adjustment factor is 0.85, and the adaptive threshold is 22.7; in the working condition of high flow and large temperature difference, the threshold adjustment factor is 1.25, and the adaptive threshold is 33.4. This adaptive mechanism can adapt to the change of the system fluctuation characteristics under different working conditions and reduce false alarms and missed alarms.

[0195] In the process of constructing the historical distribution feature vector, the quality of data preprocessing directly affects the accuracy of subsequent analysis. The method of detecting outliers can use statistical methods such as the 3σ criterion or the box plot method. For example, for the fault feature index of a certain double bypass valve, the mean is 15.3 and the standard deviation is 3.8. Using the 3σ criterion, data points less than 4.0 or greater than 26.6 are determined as outliers. In the actual data, 15 outliers are detected, accounting for 0.14% of the total data. These outliers are caused by sensor failure or system transient impact and should be excluded in subsequent analysis. Data normalization can use minimum-maximum normalization or Z-score normalization. For the statistical characteristics of the fault feature index distribution, in addition to the basic mean and standard deviation, higher-order statistics or quantiles can be calculated to more comprehensively describe the distribution characteristics. For example, the 90%, 95%, and 99% quantiles are 20.6, 22.4, and 25.1, respectively. These quantile values have important reference value for setting warning thresholds with different confidence levels.

[0196] In the process of calculating skewness and kurtosis, attention should be paid to the influence of sample size on estimation accuracy. Generally, the larger the sample size, the more accurate the estimation. For small sample sizes, bias correction techniques can be used to improve estimation accuracy. For example, for a small cooling system with only 30 days of operation data, the corrected skewness calculation formula gives a skewness of 0.22, which is more accurate than the uncorrected value of 0.25. The interpretation of skewness and kurtosis needs to be combined with specific application scenarios. For fault feature indexes, positive skewness (such as 0.18) indicates that the distribution is right-tailed, i.e., there are a small number of large index values, which is usually related to occasional small failures or unstable states of the system; negative kurtosis (such as -0.25) indicates that the distribution is flatter than the normal distribution, i.e., moderate amplitude fluctuations are more common, which is related to the diversification of system operating conditions.

[0197] Operating parameter classification is the basis for constructing operating condition feature vectors. Classification criteria should consider the physical meaning of parameters, the degree of influence on system performance, and the feasibility of monitoring. For example, the operating parameters of a certain large industrial cooling system are divided into four categories: fluid parameters, thermal parameters, mechanical parameters, and electrical parameters. Fluid parameters include main loop flow (range 0-500 m 3 / h), bypass flow (range 0-150 m 3The system pressure (range 0-1.6 MPa); thermal parameters include the inlet temperature (range 15-35 °C), the outlet temperature (range 5-25 °C), the temperature difference (range 5-15 °C); mechanical parameters include the vibration intensity (range 0-5 mm / s), the noise level (range 60-90 dB); electrical parameters include the motor current (range 20-100 A), the power factor (range 0.8-0.95). The most representative parameters from each category are selected to form the working condition feature vector, such as the main loop flow, the system pressure, the temperature difference, the vibration intensity, and the motor current. This classification method ensures that the working condition feature vector can comprehensively reflect the working state of the system.

[0198] The calculation of the correlation coefficient and the contribution coefficient needs to consider the multicollinearity problem between parameters. When there is a strong correlation between multiple working condition parameters, the directly calculated correlation coefficient and the regression coefficient are not accurate. The solution includes principal component regression or ridge regression techniques. For example, for a certain cooling system, the correlation coefficient between the inlet temperature and the outlet temperature is 0.92, indicating a strong correlation. Using the principal component regression method, the two parameters are combined into a temperature feature, and then the correlation between this feature and the fault feature index is calculated to obtain a more accurate contribution evaluation. In addition, the correlation analysis should also consider nonlinear relationships. For example, through scatter plot analysis, it is found that there is an obvious nonlinear relationship between the vibration intensity and the fault feature index, and the linear correlation coefficient is only 0.38, but the Spearman correlation coefficient based on rank is 0.65, which more accurately reflects the monotonic relationship between the two.

[0199] In the process of constructing the working condition parameter combination coefficient matrix, the identification and quantification of interaction is the key. The interaction between two parameters can be evaluated by analyzing their joint effect on the fault feature index. For example, for the flow and temperature difference parameters, when the flow is increased alone, the fault feature index increases by an average of 2.5; when the temperature difference is increased alone, the fault feature index increases by an average of 3.8; but when the flow and temperature difference are increased simultaneously, the fault feature index increases by an average of 7.2, which is greater than the sum of the individual effects, indicating a positive interaction. The interaction coefficient can be defined as the ratio of the actual joint effect to the sum of the independent effects, which is 7.2 / (2.5+3.8)=1.15 in this example. Similarly, the interaction coefficients between all pairs of parameters can be calculated to form the non-diagonal elements of the working condition parameter combination coefficient matrix, and the diagonal elements are the contribution coefficients of the parameters.

[0200] In the parameter distribution matrix construction process, the division of the operating parameter interval should consider the actual operating characteristics and data distribution of the system. Interval division can use the equal width method or the equal frequency method, or key boundary points can be determined according to the physical characteristics or operating specifications of the system. For example, for the main loop flow, according to the system design and operating experience, 0-150m³ / h can be defined as the low flow interval, 150-350m 3 / h defined as the medium flow interval, and 350-500m 3 / h defined as the high flow interval. The calculation of the conditional distribution variance requires sufficient sample size to ensure the reliability of statistical estimation. For example, for the high flow and large temperature difference operating condition combination, if there are only a small number of corresponding samples in the historical data, interpolation estimation using neighboring operating condition data can be used, or supplementary data can be generated through system simulation.

[0201] The construction of the adaptive threshold mapping function is the core of the adaptive operating condition fault warning, and the form of the mapping function can be linear or nonlinear, depending on the complexity of the correlation characteristics. A commonly used method is to construct a piecewise linear function, and different mapping relationships are used for different operating condition intervals. For example, for a certain double bypass valve, according to the combination of flow and temperature difference, the operating condition space is divided into 9 regions, and an adjustment factor for the threshold value is set for each region. In actual application, when the system operating condition changes, the fault warning threshold value is automatically adjusted. For example, when the system switches from the medium flow and medium temperature difference operating condition (threshold value of 26.7) to the high flow and large temperature difference operating condition, the warning threshold value is automatically adjusted to 33.4, which adapts to the increased volatility of the system in the latter operating condition. This adaptive mechanism significantly improves the accuracy of the warning, and in the actual application of a certain industrial cooling system, the false alarm rate is reduced from 12% to 3%, and the missed alarm rate is reduced from 8% to 2%.

[0202] The setting of the benchmark threshold needs to balance the sensitivity and specificity, a lower threshold can capture early signs of failure, but will increase the false positive rate; a higher threshold will lead to false negatives, delay failure handling. Usually, the benchmark threshold is set by adding multiple standard deviations to the mean, the multiple can be determined according to the needs of the application scenario. For example, for the key cooling system, the mean plus 2 times the standard deviation can be used as the benchmark threshold to improve the sensitivity; for the system with high fault tolerance, the mean plus 4 times the standard deviation can be used to reduce false positives. In practical applications, an industrial cooling system realizes accurate early warning of double bypass valve failure through adaptive threshold technology. After 6 months of operation, the fault feature index increased slightly from an average of 15.3 to 22.8, but was still lower than the fixed threshold of 26.7, which could not be identified by the traditional method; while the adaptive threshold system considered the low flow and small temperature difference at that time, the warning threshold was 22.7, successfully triggering the warning. Maintenance personnel found that the first bypass valve control signal line was not in good contact, and timely processing avoided subsequent more serious failures.

[0203] In an optional implementation, the parameter distribution matrix is coupled and mapped with the working condition parameter combination coefficient matrix to extract the correlation characteristics of the fault feature index and the working condition parameters, an adaptive threshold mapping function is constructed based on the correlation characteristics, and a fault warning threshold of the fault feature index is determined according to the adaptive threshold mapping function, including:

[0204] A distribution density function and a cumulative distribution function of the fault feature index are calculated based on the parameter distribution matrix to construct a probability distribution feature vector of the fault feature index;

[0205] Mutual information and conditional entropy between the working condition parameters are calculated using the working condition parameter combination coefficient matrix to construct a state transition matrix of the working condition characteristics;

[0206] The probability distribution feature vector is mapped and operated with the state transition matrix, and a time-varying correlation degree of the fault feature index and the working condition parameters is calculated as the correlation characteristics based on the mutual information and the conditional entropy;

[0207] A dynamic weight matrix of the working condition parameters is constructed according to the time-varying correlation degree, a comprehensive influence weight of the working condition parameters is calculated using the dynamic weight matrix, and a multi-dimensional mapping space of the working condition characteristics is constructed based on the comprehensive influence weight;

[0208] An adaptive threshold mapping function considering working condition conversion is established in the multi-dimensional mapping space, the benchmark threshold of the working condition steady state interval and the dynamic compensation factor of the working condition conversion interval in the adaptive threshold mapping function are adaptively fused to determine the fault warning threshold of the fault feature index.

[0209] The parameter distribution matrix is coupled and mapped with the working condition parameter combination coefficient matrix to extract the correlation characteristics of the fault feature index and the working condition parameters, and an adaptive threshold mapping function is constructed based on the correlation characteristics, and a fault warning threshold of the fault feature index is determined according to the adaptive threshold mapping function. The parameter distribution matrix contains the statistical distribution characteristics of the fault feature index under different working conditions, the working condition parameter combination coefficient matrix reflects the interaction relationship between the working condition parameters, the coupling mapping of the two can reveal the internal relationship between the fault feature index and the working condition parameters, and lays a foundation for constructing the adaptive threshold mapping function.

[0210] The distribution density function and the cumulative distribution function of the fault feature index are calculated based on the parameter distribution matrix, and a probability distribution feature vector of the fault feature index is constructed. For the fault feature index data under each condition in the parameter distribution matrix, a probability density function is constructed by using a kernel density estimation method; a Gaussian kernel is selected as the kernel function, and a bandwidth parameter is determined by a cross-validation method. The probability density function is integrated to obtain a cumulative distribution function; in actual calculation, the numerical integration of the probability density function or the method of the empirical distribution function can be used. Feature parameters of the distribution density function are extracted, such as mode, mean, variance, skewness, kurtosis, etc.; feature parameters of the cumulative distribution function are extracted, such as quantile values; and the feature parameters are combined into a vector form.

[0211] For example, for a double bypass valve of an industrial cooling system, under the working condition of a flow rate of 300 m 3 / h and a temperature difference of 10℃, the probability density function of the fault feature index presents a unimodal characteristic, the mode is 16.5, the mean is 17.2, the variance is 15.8, the skewness is 0.25, and the kurtosis is -0.18; the 10%, 25%, 50%, 75%, and 90% quantiles of the cumulative distribution function are 12.4, 14.6, 17.2, 19.8, and 22.3, respectively. These parameters constitute a probability distribution feature vector, which comprehensively describes the distribution characteristics of the fault feature index under this working condition.

[0212] The mutual information and the conditional entropy between the working condition parameters are calculated by using the working condition parameter combination coefficient matrix, and a state transition matrix of the working condition feature is constructed. For each pair of parameters in the working condition parameter combination coefficient matrix, the joint probability distribution and the marginal probability distribution are estimated according to historical data; the mutual information is calculated based on the distributions, and the greater the mutual information, the stronger the dependence between the two parameters. For each pair of parameters in the working condition parameter combination coefficient matrix, the conditional entropy is calculated based on the joint probability distribution and the marginal probability distribution; the smaller the conditional entropy, the lower the uncertainty of the other parameter when one parameter is known. The time sequence variation mode of the working condition parameters is analyzed; the transition frequencies between different working condition states are counted; the transition frequencies are normalized to obtain state transition probabilities; and the state transition probabilities are organized into a matrix form.

[0213] For example, for an industrial cooling system including three main parameters of flow rate, pressure and temperature difference, the mutual information between flow rate and pressure is 0.62, the mutual information between flow rate and temperature difference is 0.45, and the mutual information between pressure and temperature difference is 0.38; under the condition of known flow rate, the conditional entropy of pressure is 0.72, and the conditional entropy of temperature difference is 0.85. By analyzing the historical change data of the three parameters, an 8x8 state transition matrix (each parameter is divided into low and high states, a total of 2 3 =8 combined states) is constructed. The elements in the matrix represent the probability of the system transitioning from one working condition state to another, for example, the probability of transitioning from the "low flow rate, low pressure, low temperature difference" state to the "high flow rate, low pressure, low temperature difference" state is 0.18.

[0214] The probability distribution feature vector is mapped with the state transition matrix, and the time-varying correlation degree of the fault feature index and the working condition parameters is calculated based on the mutual information and the conditional entropy as the correlation characteristics. For each working condition transition pair in the state transition matrix, the corresponding probability distribution feature vector is extracted; the similarity between the two distribution feature vectors is calculated, which can be measured by cosine similarity or Kullback-Leibler divergence; the correlation strength of the working condition transition pair is calculated according to the similarity and the transition probability. Combined with the mutual information and the conditional entropy, a parameter importance evaluation function is constructed; for each working condition parameter, the contribution of its to the change of the fault feature index distribution is calculated; considering the interaction between parameters, the comprehensive correlation degree is calculated; the dynamic characteristics of the correlation degree with the change of working condition are analyzed, and the time-varying correlation degree function is obtained.

[0215] For example, for the above industrial cooling system, it is found that the time-varying correlation degree of the flow rate parameter fluctuates between 0.65-0.82, when the system switches from low flow rate to high flow rate, the correlation degree rapidly rises to 0.82, and then slowly decreases to 0.75 after stable operation for a period of time in the high flow rate state; the time-varying correlation degree of the temperature difference parameter changes between 0.55-0.78, reaching a maximum of 0.78 when the system load changes rapidly; the time-varying correlation degree of the pressure parameter is relatively stable, maintaining between 0.58-0.65. This time-varying correlation degree analysis reveals the dynamic change law of the correlation strength between the fault feature index and the working condition parameters, providing a basis for adaptive threshold adjustment.

[0216] A dynamic weight matrix of the working condition parameters is constructed according to the time-varying correlation degree, a comprehensive influence weight of the working condition parameters is calculated by using the dynamic weight matrix, and a multi-dimensional mapping space of the working condition characteristics is constructed based on the comprehensive influence weight. Based on the time-varying correlation degree, a weight coefficient of each working condition parameter in different working condition states and transition processes is calculated; the weight coefficients are organized into a matrix form, with rows representing different working condition states and columns representing different working condition parameters. By considering the interaction between the parameters, the dynamic weight matrix is converted into a comprehensive influence weight by a weighted combination or a nonlinear mapping method; the comprehensive influence weight not only considers the importance of a single parameter, but also considers the synergistic effect of the parameter combination.

[0217] A multi-dimensional space is established with the working condition parameters as coordinate axes; a comprehensive influence weight value is associated with each space point; and a continuous weight distribution function is constructed by interpolation or fitting method. For example, for three parameters of flow rate, temperature difference and pressure, a 3-dimensional working condition characteristic mapping space is constructed. In the space, each point corresponds to a working condition state and is associated with a comprehensive influence weight value. At a working condition point of a flow rate of 350 m 3 / h, a temperature difference of 12 ℃ and a pressure of 1.2 MPa, the comprehensive influence weight is 0.82; and at a working condition point of a flow rate of 150 m 3 / h, a temperature difference of 7 ℃ and a pressure of 0.8 MPa, the comprehensive influence weight is 0.65. The multi-dimensional mapping space directly shows the influence degree of different working conditions on the fault feature index, and provides a spatial framework for constructing an adaptive threshold mapping function.

[0218] An adaptive threshold mapping function considering working condition conversion is established in the multi-dimensional mapping space, the reference threshold of the working condition steady state interval and the dynamic compensation factor of the working condition transition interval in the adaptive threshold mapping function are adaptively fused, and a fault warning threshold of the fault feature index is determined. In the multi-dimensional mapping space, the working condition states are divided into a steady state interval and a transition interval; for the steady state interval, a reference threshold is set based on the probability distribution characteristics of the fault feature index, and is usually a high quantile value, such as a 95% quantile; for the transition interval, a dynamic compensation factor is calculated, considering the influence of working condition change on the fluctuation of the fault feature index.

[0219] According to the current working condition state and the change trend, the type of the interval in which the system is located is determined; for the steady state interval, the reference threshold is directly used; for the transition interval, the reference threshold is combined with the dynamic compensation factor to calculate an adjusted threshold; the rate and amplitude of the working condition change are considered to optimize the response characteristics of the threshold adjustment. The changes of the working condition parameters are monitored in real time; the current working condition is mapped to the corresponding point in the multi-dimensional space; the fault warning threshold under the current working condition is calculated by the adaptive threshold mapping function; and the fault feature index is compared with the warning threshold to determine whether the warning is triggered. For example, an industrial cooling system is in a stable running state (flow rate of 350 m3 The 95% quantile of the fault feature index is 24.8 under the condition of 200 m

[0220] When the system starts to switch to another working condition (flow rate 200 m 3 When the system starts to switch to another working condition (flow rate 200 m

[0221] The calculation of the distribution density function and the cumulative distribution function is crucial for accurately describing the statistical characteristics of the fault feature index. In the kernel density estimation method, the selection of the kernel function and the determination of the bandwidth parameter directly affect the estimation accuracy. Commonly used kernel functions include Gaussian kernel, Epanechnikov kernel, and triangular kernel, etc. Gaussian kernel has good smoothing and mathematical properties, and is the most commonly used choice. The determination of the bandwidth parameter can use Silverman rule or cross-validation method. For example, for the fault feature index data of a certain double bypass valve (sample size is 500), using Gaussian kernel function, the optimal bandwidth parameter determined by cross-validation method is 1.35.

[0222] Using this parameter for kernel density estimation, the probability density function obtained is smooth and continuous in the main peak area (15-20), and can accurately capture the distribution characteristics of the data in the tail area (>25). The calculation of the cumulative distribution function can be realized by numerical integration of the probability density function, or can be directly derived from the empirical distribution function. For example, for the above data, the probability of the fault feature index being greater than 25 is 0.058, and the probability of being greater than 30 is 0.012, which has important reference significance for setting a reasonable early warning threshold.

[0223] The calculation of mutual information and conditional entropy involves the estimation of probability distribution. For continuous working condition parameters, kernel density estimation or histogram method can be used to estimate the probability distribution; for discrete parameters, the frequency can be directly counted. When estimating the joint probability distribution, the curse of dimensionality problem needs to be considered, and when the number of parameters increases, the required sample size increases exponentially. A solution is to use a parametric model such as Gaussian mixture model to reduce the number of parameters to be estimated.

[0224] For example, for both flow rate and pressure, a bivariate Gaussian mixture model (with 3 components) is used to fit their joint distribution, and the model parameters are estimated by the expectation-maximization algorithm. Based on this model, the calculated mutual information is 0.62, which is close to the result of the non-parametric method (0.65), but is more computationally efficient. The calculation of conditional entropy can also be based on a parametric model or a non-parametric method. For example, the conditional entropy of pressure given the flow rate is 0.72, indicating that even if the flow rate value is known, there is still considerable uncertainty in the pressure, which reflects the complex relationship between the two parameters.

[0225] The construction of the state transition matrix requires the analysis of the time series pattern of the operating parameters. For continuously operating industrial systems, the operating states usually follow certain transition rules, which can be captured by Markov models or hidden Markov models. When constructing the state transition matrix, an appropriate time scale needs to be set. A too short time scale will result in noise interference, and a too long time scale will lose important dynamic characteristics.

[0226] For example, for a certain cooling system, 30 days of operating data are analyzed with a time resolution of 1 hour. By discretizing the operating parameters (each parameter is divided into low and high states), 8 combined states are obtained. The state transition frequencies of adjacent time points are counted to construct an 8x8 state transition matrix. Matrix analysis shows that the system has a clear state transition tendency, such as from the "low flow, low pressure, low temperature difference" state to the "low flow, high pressure, low temperature difference" state (probability of 0.32) rather than the "high flow, low pressure, high temperature difference" state (probability of 0.05). This transition characteristic reflects the physical constraints and operating rules of the system.

[0227] The calculation of time-varying correlation needs to consider mutual information, conditional entropy, and operating transition characteristics. One method is to construct an information gain-based parameter importance evaluation function, which measures the contribution of a parameter to reducing the uncertainty of the fault feature index. For example, the information gain of the flow rate parameter is 0.45, indicating that knowing the flow rate value can reduce the uncertainty of the fault feature index by 45%. Time-varying correlation also needs to consider the dynamic impact of operating changes. Generally, when the operating conditions change rapidly, the correlation will significantly increase. For example, when the system suddenly switches from low flow to high flow, the time-varying correlation of the flow rate parameter increases from 0.65 to 0.82, reflecting the strong impact of flow rate changes on the fault feature index during the operating transition process. This dynamic change characteristic is crucial for constructing an accurate adaptive threshold mapping function.

[0228] The calculation of dynamic weight matrix and comprehensive influence weight is the basis for constructing multi-dimensional mapping space. The dynamic weight matrix not only considers the importance of parameters in steady-state conditions, but also considers the influence of parameters in the process of working condition conversion. For example, the weight of temperature difference parameter in steady-state condition is 0.65, but the weight rises to 0.78 when the load changes rapidly, indicating that the influence of temperature difference change on fault feature index is enhanced in the dynamic process. The calculation of comprehensive influence weight needs to consider the interaction between parameters, and simple linear weighting cannot accurately reflect the complex parameter interaction. Nonlinear combination based on Choquet integral is adopted, which can handle the synergistic effect and redundancy effect between parameters. For example, the weights of flow and pressure are 0.75 and 0.65 respectively, but the combined weight of them is 0.85, which is less than the sum of the two, indicating that there is certain information redundancy; while the combined weight of flow and temperature difference is 1.45, which is greater than the sum of the two, indicating that there is synergistic enhancement effect.

[0229] An adaptive threshold mapping function is constructed, and the function design needs to consider the different characteristics of steady-state interval and conversion interval. For steady-state interval, the reference threshold can be set based on the high quantile of fault feature index, such as 95% or 99% quantile, the specific selection depends on the tolerance of false alarm rate and false alarm rate in application scenario. For example, for critical cooling system, 95% quantile can be selected as reference threshold to improve the sensitivity to abnormality; while for general system, 99% quantile can be selected to reduce false alarm. For conversion interval, the calculation of dynamic compensation factor needs to consider the rate, amplitude and direction of working condition change.

[0230] For example, when the system rapidly switches from low load to high load, the compensation factor needs to be set to 1.2-1.5 to adapt to the increase of system volatility; while when the system gradually decreases from high load to low load, the compensation factor is between 1.0-1.2. In the adaptive fusion process, a smooth transition function can be used to avoid the mutation of threshold. For example, sigmodi function is used to realize the smooth transition from steady-state threshold to conversion threshold, and the transition rate parameter can be dynamically adjusted according to the rate of working condition change.

[0231] In practical applications, the performance of the adaptive threshold mapping function needs to be verified by historical data. For example, an industrial cooling system records all fault events and corresponding fault characteristic index values, as well as changes in operating parameters during a year of operation. Through backtesting analysis, the false positive rate of the traditional fixed threshold method is 15%, and the false negative rate is 8%; after using the adaptive threshold method, the false positive rate is reduced to 3%, and the false negative rate is reduced to 2%, and the overall early warning accuracy is increased from 80% to 95%. Especially during periods of frequent operating conditions, the adaptive method has a more obvious advantage, with an early warning accuracy that is 25% higher than the fixed threshold method. These performance improvements directly translate into reduced maintenance costs and improved system reliability, for example, the annual maintenance cost of a certain system is reduced by about 32%, and the unplanned downtime is reduced by about 45%.

[0232] The industrial cooler double bypass valve fault diagnosis and early warning system of the embodiment of the application comprises:

[0233] The first unit is configured to obtain real-time operating parameters of the industrial cooler double bypass valve, and construct an operating state feature vector of the industrial cooler double bypass valve based on the real-time operating parameters;

[0234] The second unit is configured to perform dynamic time series analysis on the operating state feature vector, calculate a sliding entropy value of the operating state feature vector and a Lyapunov exponent of a reconstructed trajectory in a phase space, and determine a state evolution trend feature of the industrial cooler double bypass valve according to the sliding entropy value and the Lyapunov exponent;

[0235] The third unit is configured to calculate a fault characteristic index of the industrial cooler double bypass valve according to the state evolution trend feature and a deviation value of a valve opening degree and an input control signal in the operating state feature vector;

[0236] The fourth unit is configured to determine a fault early warning threshold of the fault characteristic index based on a historical data distribution variance of the fault characteristic index and in combination with an operating condition parameter of the industrial cooler;

[0237] The fifth unit is configured to determine a fault type of the industrial cooler double bypass valve according to a feature combination of the fault characteristic index that exceeds the fault early warning threshold, and output corresponding fault early warning information when the fault characteristic index exceeds the fault early warning threshold.

[0238] In a third aspect, the embodiment of the application provides an electronic device, comprising:

[0239] a processor;

[0240] a memory for storing processor-executable instructions;

[0241] The processor is configured to invoke the instructions stored in the memory to execute the method described above.

[0242] In a fourth aspect, the present application provides a computer readable storage medium, having stored thereon computer program instructions, which when executed by a processor, implement the method described above.

[0243] The present application can be a method, an apparatus, a system, and / or a computer program product. The computer program product can include a computer readable storage medium (or media) having computer readable program instructions thereon for performing various aspects of the present application.

[0244] Finally, it should be noted that: the above embodiments are only used to illustrate the technical solutions of the present application, and not to limit them; although the present application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand: it can still modify the technical solutions recorded in the foregoing embodiments, or make equivalent replacement for part or all of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the scope of the technical solutions of the embodiments of the present application.

Claims

1. An industrial chiller dual bypass valve fault diagnosis and early warning method, characterized in that, The method comprises the following steps: acquiring real-time operation parameters of the industrial cooler double bypass valve, and constructing an operation state feature vector of the industrial cooler double bypass valve based on the real-time operation parameters; performing dynamic time series analysis on the operation state feature vector, calculating a sliding entropy value of the operation state feature vector and a Lyapunov exponent of a phase space reconstruction trajectory, and determining a state evolution trend feature of the industrial cooler double bypass valve according to the sliding entropy value and the Lyapunov exponent; calculating a fault feature index of the industrial cooler double bypass valve according to the state evolution trend feature and a deviation value of a valve opening degree and an input control signal in the operation state feature vector; determining a fault warning threshold of the fault feature index based on a historical data distribution variance of the fault feature index and in combination with an operation condition parameter of the industrial cooler; when the fault feature index exceeds the fault warning threshold, determining a fault type of the industrial cooler double bypass valve according to a feature combination of the fault feature index exceeding the threshold, and outputting corresponding fault warning information; calculating a fault feature index of the industrial cooler double bypass valve according to the state evolution trend feature and a deviation value of a valve opening degree and an input control signal in the operation state feature vector, comprising: constructing a dynamic transfer matrix based on the state evolution trend feature, and constructing a valve response feature matrix based on the deviation value of the valve opening degree and the input control signal in the operation state feature vector; performing collaborative mapping on the dynamic transfer matrix and the valve response feature matrix to obtain a dynamic and static feature transfer path of the double bypass valve and a valve response topological correlation degree; extracting valve opening degree distribution characteristics of the double bypass valve in different operation states according to the dynamic and static feature transfer path, and constructing a spatial mapping matrix of the valve opening degree; calculating an opening response correlation strength and an opening synchronicity index of the double bypass valve based on the valve response topological correlation degree, and constructing a correlation mapping matrix of the valve opening degree; performing feature fusion on the spatial mapping matrix of the valve opening degree and the correlation mapping matrix of the valve opening degree to obtain a valve opening degree comprehensive evaluation index of the double bypass valve, and calculating a fault feature index of the industrial cooler double bypass valve according to the valve opening degree comprehensive evaluation index.

2. The method of claim 1, wherein, The method comprises the following steps: acquiring real-time operation parameters of the industrial cooler double bypass valve, and constructing an operation state feature vector of the industrial cooler double bypass valve based on the real-time operation parameters; acquiring real-time operation parameters of a first bypass valve and a second bypass valve of the industrial cooler double bypass valve during operation; performing time domain analysis on the real-time operation parameters, extracting time domain feature parameters of the first bypass valve and the second bypass valve, and performing wavelet decomposition on the time domain feature parameters to obtain high-frequency components and low-frequency components of the first bypass valve and the second bypass valve; extracting transient response features according to the high-frequency components, extracting steady-state operation features according to the low-frequency components, combining the transient response features and the steady-state operation features, and obtaining dynamic and static feature parameters of the double bypass valve; According to the dynamic and static characteristic parameters, a state space matrix of the first bypass valve and the second bypass valve is constructed, and a main eigenvector is determined based on a singular value decomposition result of the state space matrix; The main eigenvector is projected into a preset characteristic space to obtain a characteristic mapping matrix, and the operating state characteristic vector is extracted according to the characteristic mapping matrix.

3. The method of claim 1, wherein, Dynamic time series analysis is performed on the operating state characteristic vector, a sliding entropy value of the operating state characteristic vector and a Lyapunov exponent of a phase space reconstruction trajectory are calculated, and a state evolution trend feature of the industrial cooler double bypass valve is determined according to the sliding entropy value and the Lyapunov exponent, including: The operating state characteristic vector is divided into a plurality of time windows according to a time sequence, and the sliding entropy value of the operating state characteristic vector is calculated in each time window; A phase space reconstruction trajectory is constructed based on the operating state characteristic vector, and an embedding dimension and a time delay parameter of the phase space reconstruction trajectory are determined; The phase space trajectory of the operating state characteristic vector is reconstructed according to the embedding dimension and the time delay parameter, and the Lyapunov exponent of the phase space trajectory is calculated; Multi-scale fusion analysis is performed on the sliding entropy value and the Lyapunov exponent, and a cooperative evolution matrix of the double bypass valve state feature is constructed; Feature decomposition is performed on the cooperative evolution matrix to obtain a feature value sequence, a dominant eigenvalue corresponding to the largest amplitude is obtained by sorting the feature value sequence according to the amplitude, and a state evolution characteristic space is constructed based on the dominant eigenvalue; The convergence radius and the divergence coefficient of the state trajectory are calculated in the state evolution characteristic space, and the state evolution trend feature of the industrial cooler double bypass valve is determined according to the convergence radius and the divergence coefficient.

4. The method of claim 3, wherein, Multi-scale fusion analysis is performed on the sliding entropy value and the Lyapunov exponent, and a cooperative evolution matrix of the double bypass valve state feature is constructed, including: A bivariate cooperative matrix is constructed according to the sliding entropy value and the Lyapunov exponent, recursive quantization analysis is performed on the bivariate cooperative matrix, and a dynamic association network of the state feature is established; Topological features and transmission characteristics of the dynamic association network are extracted, and a coupling relationship between the state features is determined according to the topological features and the transmission characteristics; Wavelet transform is performed on the coupling relationship to obtain wavelet coefficients at different scales, and the wavelet coefficients are divided into a high-frequency subband and a low-frequency subband according to an energy distribution law; A multi-scale feature matrix is constructed based on the high-frequency subband and the low-frequency subband, and a principal component eigenvector is obtained by performing singular value decomposition on the multi-scale feature matrix; The principal component eigenvector is projected into a preset state evolution space to obtain a state evolution projection matrix, and the cooperative evolution matrix of the double bypass valve state feature is constructed according to the state evolution projection matrix.

5. The method of claim 1, wherein, Based on the distribution variance of the historical data of the fault feature index and combined with the operating condition parameters of the industrial cooler, a fault warning threshold of the fault feature index is determined, including: Distribution characteristic analysis is performed on the fault feature index to construct a historical distribution characteristic vector of the fault feature index; According to the historical distribution feature vector, a distribution skewness and a distribution kurtosis of a fault feature index are calculated, and a historical data distribution variance of the fault feature index is obtained; The operating condition parameters are classified according to the condition types, a condition feature vector is constructed, and a correlation coefficient and a contribution degree coefficient of each parameter in the condition feature vector are calculated; A condition parameter combination coefficient matrix is constructed based on the correlation coefficient and the contribution degree coefficient, and a parameter distribution matrix is constructed based on the historical data distribution variance; The parameter distribution matrix and the condition parameter combination coefficient matrix are coupled and mapped, the associated characteristics of the fault feature index and the condition parameters are extracted, an adaptive threshold mapping function is constructed based on the associated characteristics, and a fault warning threshold of the fault feature index is determined according to the adaptive threshold mapping function.

6. The method of claim 5, wherein, The parameter distribution matrix and the condition parameter combination coefficient matrix are coupled and mapped, the associated characteristics of the fault feature index and the condition parameters are extracted, an adaptive threshold mapping function is constructed based on the associated characteristics, and a fault warning threshold of the fault feature index is determined according to the adaptive threshold mapping function, including: A distribution density function and a cumulative distribution function of the fault feature index are calculated based on the parameter distribution matrix, and a probability distribution feature vector of the fault feature index is constructed; The mutual information and the conditional entropy between the condition parameters are calculated using the condition parameter combination coefficient matrix, and a state transition matrix of the condition feature is constructed; The probability distribution feature vector and the state transition matrix are mapped and operated, the time-varying correlation degree of the fault feature index and the condition parameters is calculated based on the mutual information and the conditional entropy as the associated characteristics; A dynamic weight matrix of the condition parameters is constructed according to the time-varying correlation degree, the comprehensive influence weight of the condition parameters is calculated using the dynamic weight matrix, and a multi-dimensional mapping space of the condition feature is constructed based on the comprehensive influence weight; An adaptive threshold mapping function considering condition conversion is established in the multi-dimensional mapping space, the reference threshold of the condition steady state interval and the dynamic compensation factor of the condition conversion interval in the adaptive threshold mapping function are adaptively fused, and the fault warning threshold of the fault feature index is determined.

7. An industrial chiller dual bypass valve fault diagnostic warning for implementing the method of any of claims 1-6, characterized by, Including: A first unit is configured to acquire real-time operating parameters of an industrial cooler double bypass valve, and construct an operating state feature vector of the industrial cooler double bypass valve based on the real-time operating parameters; A second unit is configured to perform dynamic time series analysis on the operating state feature vector, calculate a sliding entropy value of the operating state feature vector and a Lyapunov exponent of a phase space reconstruction trajectory, and determine a state evolution trend feature of the industrial cooler double bypass valve according to the sliding entropy value and the Lyapunov exponent; A third unit is configured to calculate a fault feature index of the industrial cooler double bypass valve according to the state evolution trend feature and a deviation value of a valve opening degree and an input control signal in the operating state feature vector; A fourth unit is configured to determine a fault warning threshold of the fault feature index based on a historical data distribution variance of the fault feature index and in combination with operating condition parameters of the industrial cooler. The fifth unit is configured to determine a fault type of the double bypass valve of the industrial cooler according to a characteristic combination of the fault characteristic index exceeding the threshold when the fault characteristic index exceeds the fault warning threshold, and output corresponding fault warning information. The third unit is configured to: construct a dynamic transfer matrix based on the state evolution trend characteristics, and construct a valve response feature matrix based on a deviation value of the valve opening in the operating state feature vector and the input control signal; perform collaborative mapping of the dynamic transfer matrix and the valve response feature matrix to obtain a dynamic-static feature transfer path of the double bypass valve and a valve response topology correlation degree; extract valve opening distribution characteristics of the double bypass valve in different operating states according to the dynamic-static feature transfer path, and construct a spatial mapping matrix of the valve opening; calculate an opening response correlation strength and an opening synchronism index of the double bypass valve based on the valve response topology correlation degree, and construct a correlation mapping matrix of the valve opening; perform feature fusion of the spatial mapping matrix of the valve opening and the correlation mapping matrix of the valve opening to obtain a valve opening comprehensive evaluation index of the double bypass valve, and calculate a fault characteristic index of the double bypass valve of the industrial cooler according to the valve opening comprehensive evaluation index.

8. An electronic device, comprising: comprise: a processor; a memory for storing processor-executable instructions; wherein the processor is configured to invoke the instructions stored in the memory to execute the method of any one of claims 1 to 6.

9. A computer-readable storage medium having stored thereon computer program instructions, wherein, The computer program instructions, when executed by the processor, implement the method of any one of claims 1 to 6.

Images