A method, system, device, and storage medium for fault detection of a pneumatic control valve.

Abstract

This invention provides a fault detection method, system, device, and storage medium for pneumatic control valves, belonging to the field of industrial process fault detection. The method includes: introducing a nonlinear modeling approach based on Gaussian process regression (GPR) into the fast layer to construct the flow coefficient C. V A high-precision nonlinear model is used, whose residual statistics can sensitively capture transient anomalies in control valve performance that deviate from expectations. In the slow layer, Activity-KSFA, a kernel-based slow feature analysis, is employed to effectively extract slowly varying features during system degradation, and T² statistics are used for slow anomaly detection. Finally, a dual-threshold and continuous decision mechanism is combined to suppress false alarms while maintaining high sensitivity. Simulation results on the DAMADICS platform show that C... V While achieving zero false alarms, the residual maintains a detection rate of over 98% for most typical faults, complementing the T² statistic and significantly enhancing the adaptability of this invention to different fault modes.

Description

Technical Field

[0001] This invention belongs to the field of industrial process fault detection, specifically relating to a fault detection method, system, equipment, and storage medium for pneumatic control valves. Background Technology

[0002] Pneumatic control valves, as terminal actuators in industrial process control systems, are widely used in core fields such as petrochemicals, power energy, and metallurgical manufacturing, undertaking critical tasks such as flow regulation, pressure control, and system protection. Research on online condition monitoring and fault detection of pneumatic control valves is of significant theoretical and engineering application value for ensuring the safe operation of industrial systems, reducing unplanned downtime losses, and achieving predictive maintenance.

[0003] Kernel Slow Feature Analysis (KSFA) maps raw data to a high-dimensional feature space by introducing kernel tricks, and performs slow feature extraction in the kernel space, effectively overcoming the limitations of linear SFA. However, existing KSFA methods still have certain empirical dependencies in kernel function selection and feature dimension determination, and a single monitoring statistic is difficult to simultaneously meet the detection needs of both rapid mutation faults and slow evolution faults.

[0004] At the mechanistic level, the flow coefficient (C) V C is a key indicator for measuring the regulating performance of a control valve, and its value directly reflects the valve's flow capacity at a specific opening degree. From a fault mechanism perspective, when a control valve experiences faults such as external leakage or valve core wear, the actual C... V The value will deviate significantly from the nominal value, leading to decreased control accuracy and deterioration of system performance. Therefore, based on C... V Residual analysis of coefficients provides an effective method for detecting control valve faults, but C V The coefficients have complex nonlinear mapping relationships with multiple variables such as valve opening, inlet and outlet pressure difference, and flow rate. Traditional linear regression or polynomial fitting methods are difficult to accurately describe this nonlinear characteristic, resulting in a large number of modeling errors mixed in the residual signal, which affects the accuracy of fault detection. Summary of the Invention

[0005] To address the aforementioned problems, this invention provides a method, system, device, and storage medium for fault detection of pneumatic control valves.

[0006] To achieve the above objectives, the present invention provides a fault detection method for a pneumatic control valve, comprising: The physical monitoring variables of the target pneumatic control valve under real-time operating conditions are collected. The physical monitoring variables include valve stem displacement, valve inlet pressure, valve outlet pressure, and fluid volume flow rate.

[0007] Based on the physical monitoring variables, the actual observed flow coefficient of the target pneumatic control valve is calculated according to the flow coefficient formula; the valve stem displacement is input into a pre-trained Gaussian process regression (GPR) model, and the predicted observed flow coefficient is output; the error value between the actual observed flow coefficient and the predicted observed flow coefficient is calculated, the residual value is compared with the preset fast layer control limit, a binary fast anomaly preliminary flag is generated, and a fast layer anomaly alarm signal is generated based on the fast anomaly preliminary flag.

[0008] The standard deviation of the physical monitoring variables is calculated using a fixed sliding window. The physical monitoring variables at each time point are concatenated with the standard deviations at the corresponding time points to form a multidimensional enhanced feature vector. The multidimensional enhanced feature vector is then input into a pre-trained KSFA model, which outputs the Hotelling T² statistic. Based on the T² statistic, an abnormal alarm signal for the slow layer is generated.

[0009] Based on the high-speed layer abnormality alarm signal and the slow-speed layer abnormality alarm signal, the fault detection of the target pneumatic control valve is realized.

[0010] Preferably, the step of generating a fast-layer anomaly alarm signal based on the initial fast anomaly flag specifically includes: generating a fast-layer anomaly alarm signal when the initial fast anomaly flag continuously reaches or exceeds a first preset threshold; the step of generating a slow-layer anomaly alarm signal based on the T² statistic specifically includes: setting the initial anomaly flag to 1 if the T² statistic exceeds a preset slow-layer control limit; and generating a slow-layer anomaly alarm signal when the anomaly flag continuously reaches or exceeds a second preset threshold.

[0011] Preferably, the Gaussian process regression (GPR) model is trained using a Gaussian process regression algorithm with a composite kernel function, and the kernel function parameters are optimized by maximizing the marginal likelihood function to obtain the GPR model.

[0012] Preferably, during training, the KSFA model calculates the RBF kernel matrix of the multidimensional augmented feature vectors of the training samples, centers the RBF kernel matrix, and calculates the time difference kernel matrix based on the centered RBF kernel matrix; calculates the generalized eigenvalues ​​of the time difference kernel matrix and the centered RBF kernel matrix; selects the sample vectors corresponding to the k smallest generalized eigenvalues ​​to form a slow feature extraction projection matrix, and obtains the Hotelling T² statistic based on the slow feature extraction projection matrix to complete the KSFA model training.

[0013] Preferably, based on the rapid layer abnormal alarm signal and the slow layer abnormal alarm signal, the fault detection of the target pneumatic control valve is realized, specifically including: the rapid layer abnormal alarm signal is used to detect transient sudden faults, the slow layer abnormal alarm signal is used to detect gradual degradation faults, and the triggering of any alarm signal determines that the target pneumatic control valve has a fault.

[0014] Preferably, the control limit of the fast layer is set based on the 3σ principle, and the first preset threshold is 5; the control limit of the slow layer is determined by a hybrid threshold strategy, which is set by the mean and standard deviation of the T² statistic of the training set, and the second preset threshold is 15.

[0015] Preferably, the step of calculating the standard deviation of the physical monitoring variables using a fixed sliding window and concatenating the physical monitoring variables at each time point with the standard deviations at the corresponding time point into a multi-dimensional enhanced feature vector specifically includes: obtaining the valve pressure difference between the valve inlet pressure and the valve outlet pressure, calculating the standard deviations of the valve stem displacement, valve pressure difference, and fluid volumetric flow rate respectively, and concatenating the valve stem displacement, valve pressure difference, and fluid volumetric flow rate at each time point with the standard deviations at the corresponding time point into a 6-dimensional enhanced feature vector.

[0016] The present invention also provides a fault detection system for a pneumatic control valve, comprising: The data acquisition module is used to collect physical monitoring variables in real time during the operation of the target pneumatic control valve. These physical monitoring variables include valve stem displacement, valve inlet pressure, valve outlet pressure, and fluid volumetric flow rate.

[0017] The calculation module is used to calculate the actual observed flow coefficient of the target pneumatic control valve based on the physical monitoring variables and the flow coefficient formula; input the valve stem displacement into a pre-trained Gaussian process regression (GPR) model and output the predicted observed flow coefficient; calculate the error value between the actual observed flow coefficient and the predicted observed flow coefficient, compare the residual value with the preset fast layer control limit, generate a binary fast anomaly preliminary flag, and generate a fast layer anomaly alarm signal based on the fast anomaly preliminary flag; calculate the standard deviation of the physical monitoring variables according to a fixed sliding window, and concatenate the physical monitoring variables at each time step with the standard deviation at the corresponding time step to form a multidimensional enhanced feature vector; input the multidimensional enhanced feature vector into a pre-trained KSFA model and output the Hotelling T² statistic, and generate a slow layer anomaly alarm signal based on the T² statistic.

[0018] The detection module is used to detect faults in the target pneumatic control valve based on the high-speed layer abnormal alarm signal and the slow-speed layer abnormal alarm signal.

[0019] The present invention also provides a computer device, including a memory, a processor, and a computer program stored in the memory, wherein the processor executes the computer program to implement any of the steps in the fault detection method for the pneumatic control valve.

[0020] The present invention also provides a computer-readable storage medium storing a computer program that, when loaded by a processor, can execute any of the steps in the fault detection method for the pneumatic control valve.

[0021] The fault detection method for a pneumatic control valve provided by this invention has the following beneficial effects: This invention uses a GPR model to fit the valve stem displacement and C V Complex nonlinear mapping of coefficients, accurate calculation of C V By constructing a fast detection layer using residuals, the problem of large modeling errors and low fault detection accuracy in traditional linear methods is solved. By processing physical monitoring variables through a fixed sliding window and constructing multi-dimensional enhanced feature vectors as input to the pre-trained KSFA model, a slow detection layer is constructed in conjunction with the Hotelling T² statistic. This weakens the empirical dependence of KSFA on kernel function selection and feature dimension determination, while achieving accurate capture of slowly evolving faults. Relying on the fusion judgment of abnormal alarm signals from the fast and slow dual layers, the limitation of a single monitoring statistic in detecting both rapid abrupt faults and slowly evolving faults is overcome, achieving full coverage detection of both types of faults and significantly improving the comprehensive fault detection capability of pneumatic control valves. Attached Figure Description

[0022] To more clearly illustrate the embodiments and design schemes of the present invention, the accompanying drawings required for this embodiment will be briefly described below. The drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0023] Figure 1 This is a flowchart illustrating a fault detection method for a pneumatic control valve according to an embodiment of the present invention; Figure 2 This is the overall framework of the mechanism residual and kernel slow feature analysis algorithm in this embodiment of the invention; Figure 3 This is a flowchart illustrating the fault detection process of the method according to an embodiment of the present invention; Figure 4 The image shows the detection results of fault F1 using the KPCA method according to an embodiment of the present invention. Figure 5 The image shows the detection results of fault F1 using the KECA method according to an embodiment of the present invention. Figure 6 The image shows the detection results of fault F1 using the MS-MWKECA method according to an embodiment of the present invention. Figure 7 The image shows the detection results of the method of the present invention for fault F1 in an embodiment of the present invention. Figure 8 The image shows the detection results of fault F5 using the KPCA method according to an embodiment of the present invention. Figure 9 The image shows the detection results of fault F5 using the KECA method according to an embodiment of the present invention. Figure 10 The image shows the detection results of fault F5 using the MS-MWKECA method according to an embodiment of the present invention. Figure 11 The image shows the detection results of the method of the present invention for fault F5 in an embodiment of the present invention; Figure 12 This is an experimental platform for an I-shaped pipeline network according to an embodiment of the present invention; Figure 13 This is the result of valve leakage fault detection in an I-shaped pipe network according to an embodiment of the present invention; Figure 14 The fault detection result is fed back by the positioner in the I-shaped pipeline network according to an embodiment of the present invention. Detailed Implementation

[0024] To enable those skilled in the art to better understand and implement the technical solutions of the present invention, the present invention will be described in detail below with reference to the accompanying drawings and specific embodiments. The following embodiments are only used to more clearly illustrate the technical solutions of the present invention and should not be construed as limiting the scope of protection of the present invention.

[0025] like Figure 2 As shown, this invention proposes a method for detecting process faults in control valves based on mechanistic residuals and kernel slow feature analysis. The overall architecture for process fault detection in pneumatic control valves based on mechanistic residuals and kernel slow feature analysis uses valve stem displacement, flow rate, inlet valve pressure, and outlet valve pressure as inputs, and generates C through a feature engineering layer. V After analyzing the flow coefficient, pressure difference ΔP, and activity characteristics, a rapid fault detection layer (nonlinear C) is used. V Modeling, C V The residual monitoring and continuous filtering capture transient anomalies, the slow fault detection layer (kernel slow feature analysis, slow feature extraction, T² statistic calculation and continuous filtering) captures slowly changing degradation features, and finally the decision fusion layer uses a two-layer fusion strategy to determine the fault. The performance is evaluated by the fault detection rate (FDR) and the false alarm rate (FAR) to achieve accurate detection of different fault modes.

[0026] Based on this, the present invention provides a fault detection method for pneumatic control valves, specifically a control valve process fault detection method based on mechanistic residual and kernel slow characteristic analysis, such as... Figure 1 As shown, it includes: S1. Real-time acquisition of physical monitoring variables of the target pneumatic control valve under operating conditions, including valve stem displacement, valve inlet pressure, valve outlet pressure, and fluid volumetric flow rate.

[0027] Bartys et al. developed the DAMADICS platform. Using three industrial actuators from the sugar refining process at the Lublin sugar factory in Poland as experimental subjects, they developed and validated the platform based on the actuators' working principles, considering their physical and electrical structural characteristics, and drawing on extensive real-world data. The signals generated by the DAMADICS simulation platform include external control signals, valve inlet pressure, valve outlet pressure, fluid medium temperature, valve stem displacement, and fluid flow rate. These signals are widely used in process fault detection and diagnosis research. The basic parameters of the benchmark platform were set based on the structural parameters of a single-seat equal-percentage pneumatic control valve, with an effective valve stem stroke of 38 mm and an effective diaphragm area of ​​290 cm². 2 The valve core diameter is 50mm, and the operating parameters include valve inlet pressure of 3.5MPa, outlet pressure of 2.6MPa, air supply pressure of 0.6MPa, and fluid medium temperature of 43℃.

[0028] This invention selects four signals generated by the DAMADICS platform—valve inlet pressure, valve outlet pressure, valve stem displacement, and fluid flow rate—as statistical monitoring variables. It collects data from the first 1900 seconds of the control valve's operation, with the fault occurrence time being the 900th second, and detects six different fault modes as shown in Table 1.

[0029] Table 1. Fault modes selected in the DAMADICS platform S2. Based on the physical monitoring variables, calculate the actual observed flow coefficient of the target pneumatic control valve according to the flow coefficient formula; input the valve stem displacement into the pre-trained Gaussian process regression (GPR) model and output the predicted observed flow coefficient; calculate the error value between the actual observed flow coefficient and the predicted observed flow coefficient, compare the residual value with the preset fast layer control limit, generate a binary fast anomaly preliminary indicator, and generate a fast layer anomaly alarm signal based on the fast anomaly preliminary indicator; calculate the standard deviation of the physical monitoring variables according to a fixed sliding window, and concatenate the physical monitoring variables at each time point with the standard deviation at the corresponding time point to form a multidimensional enhanced feature vector; input the multidimensional enhanced feature vector into the pre-trained KSFA model and output the Hotelling T² statistic, and generate a slow layer anomaly alarm signal based on the T² statistic.

[0030] Slow Feature Analysis (SFA) is an unsupervised temporal feature extraction method. Its core idea is to extract the slowest-changing latent features from high-dimensional dynamic signals. Given an input signal... for 1 variable and Given a set of samples, the SFA algorithm aims to find a set of feature functions. This makes the output signal The following optimization objectives and constraints must be met.

[0031] (1) in, The first difference of a latent variable is also called the rate of change of the latent variable. This represents the average value over time. Constraint 1 ensures that the acquired latent variables have zero mean, simplifying the problem-solving process. Constraint 2 ensures that the slow features have unit variance, guaranteeing that they carry certain information. Constraint 3 ensures that the slow features are orthogonal to each other and carry different types of information. A pneumatic control valve consists of a valve body, actuator, positioner, and other components. Its working process involves the controller outputting a 4-20mA current signal to the intelligent positioner. The positioner converts the electrical signal into a pneumatic signal to drive the diaphragm actuator, which in turn pushes the valve stem to change the valve core position, thereby regulating the fluid flow rate. Flow coefficient C V C is a core parameter characterizing the flow capacity of a valve, theoretically determined by the valve's geometry and flow characteristics. However, in actual operation, C... V The coefficient exhibits complex nonlinearity due to factors such as manufacturing tolerances, installation conditions, medium characteristics, and wear and aging. Therefore, theoretical models based on nominal flow characteristics often have significant deviations. Individualized C models are established using measured data. V A model is a prerequisite for achieving high-precision residual detection. Based on the national standard flow coefficient C... V The calculation formula is:

[0032] (2) Where Q represents volumetric flow rate, m 3 / h; This indicates the inlet pressure of the control valve, expressed in kPa. This indicates the outlet pressure of the control valve, in kPa. This invention uses water at room temperature as a medium to represent relative density. Therefore, the flow coefficient ,in This represents the pressure difference across the valve.

[0033] Mechanism residual and kernel slow characteristic analysis algorithm. Flow coefficient of equal percentage pneumatic control valve under normal operating conditions. With valve stem displacement There exists a definite nonlinear mapping relationship. When a valve malfunctions (such as external leakage, inlet leakage, jamming, etc.), this mapping relationship will deviate from the normal pattern, manifesting as... Abnormal offset of residuals. The goal of this invention is to design a detection framework to achieve the collaborative detection of fast-mutation faults and slow-degradation faults. The overall framework of the detection method proposed in this invention is as follows: Figure 1As shown, it mainly consists of three core modules: C based on Gaussian process regression. V The system includes a fast residual detection layer, a slow detection layer based on activity-enhanced kernel slow feature analysis, and a dual-threshold continuous judgment fusion mechanism.

[0034] S201, Rapid Detection Layer Due to the actual operation of C V The coefficients exhibit complex nonlinear characteristics, and the traditional C V Modeling methods struggle to accurately characterize the inherent nonlinear properties of control valves. This invention employs Gaussian process regression (GPR) to construct the relationship between valve opening and C. V A probabilistic mapping model between coefficients. A Gaussian process is a stochastic process defined on a continuous input space, whose joint distribution of any finite-dimensional component follows a multivariate Gaussian distribution. Let the objective function be... Prior to a Gaussian process:

[0035] (3) in, It is a mean function. This is the covariance kernel function. This invention uses a composite kernel function:

[0036] (4) in, For signal variance, The feature length scale is obtained automatically by maximizing the marginal likelihood function.

[0037] Given a training dataset ,in To normalize the valve opening, Given the effective sample size, the predicted distribution of GPR is as follows: (5) The predicted mean and variance are as follows: (6) (7) in, The kernel matrix of the training samples, The kernel vector between the test point and the training point. To observe the noise variance, To observe C V vector.

[0038] Based on the trained GPR model, C is calculated for any test sample. V Residual: (8) Under normal operating conditions, the residual sequence It approximately follows a zero-mean Gaussian distribution. The distribution parameters are estimated using the training set residuals.

[0039] (9) based on Principle setting C V Residual detection threshold: (10) in, For the threshold coefficient, take... .when At that time, it was determined to be a transient anomaly.

[0040] S202, Slow Detection Layer Kernel Slow Feature Analysis (KSFA). To address the inherent nonlinear characteristics of control valve systems, this invention employs the Kernel Slow Feature Analysis (KSFA) algorithm. KSFA is a nonlinear dimensionality reduction algorithm whose core idea is to extract slowly changing essential features from high-dimensional, rapidly changing observation data. Assume the input data is... The standard linear SFA seeks a linear transformation. KSFA, on the other hand, uses kernel tricks to map data to a high-dimensional feature space:

[0041] (11) Then, a linear projection is found in this space, and the objective and constraints are optimized as shown in Equation (1) to make the output signal exhibit slow characteristics. Slowest changing: (12) Due to higher-dimensional space Its dimensions may be infinitely large, making it impossible to calculate directly. According to the representation theorem, This can be represented as a linear combination of samples:

[0042] (13) in, This is the coefficient vector that needs to be solved. Therefore, slow features It can be represented in the form of a kernel function:

[0043] (14) For all training samples, written in matrix form: (15) in, It is a kernel matrix. .

[0044] Let kernel function The corresponding implicit feature mapping is ,in This invention uses a regenerable kernel Hilbert space (RKHS). It employs a radial basis function (RBF) kernel:

[0045] (16) Among them, the kernel width parameter Select adaptively based on data: (17) in, This represents the variance of the sliding window.

[0046] Given training sequence The kernel matrix is ​​defined as Its elements To eliminate the influence of the mean in the feature space, the kernel matrix is ​​centered:

[0047] (18) in, It is a vector consisting entirely of 1s.

[0048] set up and These are the centered kernel matrix blocks and the cross kernel matrix at adjacent time points. Then the time difference kernel matrix is: (19) The kernel SFA problem is transformed into the following generalized eigenvalue problem: (20) To ensure numerical stability, Perform regularization: ;(twenty one) in, For regularization parameters, take... To solve the above generalized eigenvalue problem, the smallest eigenvalue is selected. The eigenvectors corresponding to each eigenvalue Constructing a slow feature extraction matrix .

[0049] Activity feature enhancement. Early degradation in industrial processes is often accompanied by subtle changes in signal fluctuation characteristics. To capture these dynamic changes, this invention proposes an activity feature enhancement strategy. For the original feature vector... , define the first Dimensional signal in Local activity at time :

[0050] ;(twenty two) in, The mean within the window. Where is the width of the sliding window. The enhanced feature vector is:

[0051] ;(twenty three) The valve pressure difference between the valve inlet and outlet pressures is obtained. The standard deviations of valve stem displacement, valve pressure difference, and fluid volumetric flow rate are calculated separately. The valve stem displacement, valve pressure difference, and fluid volumetric flow rate at each time point are then concatenated with their corresponding standard deviations to form a 6-dimensional enhanced feature vector. This strategy increases the feature dimension from... Expand to It also integrates the instantaneous value and local fluctuation information of the signal, enhancing the sensitivity to slow degradation patterns.

[0052] T² statistic detection is based on the extracted slow features, constructing a T² statistic for process detection. Let the slow feature matrix obtained during the training phase be... Its mean vector and covariance matrix are as follows:

[0053] ;(twenty four) For the standardized slow feature vector , Define the T² statistic for the standard deviations of each dimension: (25) The control limits for the T² statistic are determined using a mixed threshold strategy: (26) in, for Quantiles, take , and These are the mean and standard deviation of the T² statistic for the training set, respectively. As adaptive coefficients, this hybrid strategy balances the characteristics of the distribution tails and the requirements of robustness.

[0054] Slow feature computation during the testing phase. For test samples. First, calculate its kernel vector relative to the training samples:

[0055] (27) The centered kernel vector is: (28) in, To train the row mean vector of the kernel matrix, This is the global mean of the training kernel matrix.

[0056] Using the slow feature extraction matrix constructed during the training phase Then the slow feature of the test sample is: (29) S3. Based on the high-speed layer abnormal alarm signal and the slow-speed layer abnormal alarm signal, the fault detection of the target pneumatic control valve is realized.

[0057] The fast and slow layers are designed for different types of failure modes, respectively. To achieve their organic integration, this invention proposes a hierarchical detection strategy:

[0058] Fast layer determination: (30) Slow layer determination: (31) in, This is an indicator function.

[0059] Continuous filtering mechanism. To suppress instantaneous false alarms caused by noise and random disturbances, this invention introduces a continuous filtering mechanism, defining a continuous anomaly counter. :

[0060] (32) A fault is only confirmed when the number of consecutive abnormal events reaches a set threshold: (33) in, and The continuity thresholds are 5 and 15 for the fast layer and the slow layer, respectively. The slow layer uses a larger continuity threshold to accommodate its detection characteristics of slow changes.

[0061] Algorithm Detection Flow. The control valve process fault detection method based on mechanistic residual and kernel slow feature analysis includes two parts: offline training and online monitoring. The complete algorithm flowchart is shown below. Figure 3 As shown. In the offline training phase, normal operating state data is collected to establish a two-layer fault detection algorithm model, and control limits for the fast detection layer and the slow detection layer are established. In the online detection phase, transition data from normal to fault states is collected, fault detection indicators are calculated and compared with the corresponding control limits to determine the operating state of the pneumatic control valve. The specific modeling steps are as follows:

[0062] S301, Offline Training Process Step 1: Collect four state variables during the normal operation of the pneumatic control valve. ; Step 2: Calculate the observed flow coefficient according to formula (2). Construct the observation vector matrix ; Step 3: Construct a GPR model based on formula (3) and fit the valve opening with C. V Mapping relationship between coefficients ; Step 4: Calculate the training set flow coefficient C using formula (8). V The residuals are estimated using formula (9). , And set the flow coefficient C according to formula (10). V control limits ; Step 5: Enhance the activity of the features according to formulas (22) and (23) to construct the feature vector. It integrates the instantaneous value and local fluctuation information of the signal; Step 6: Calculate the kernel matrix according to formula (16) The kernel matrix is ​​then centered according to formula (18). ; Step 7: Calculate the time difference kernel matrix according to formula (19) The kernel SFA problem is transformed into a generalized eigenvalue problem for solution, and the minimum eigenvalue is selected. Eigenvectors corresponding to eigenvalues Constructing the slow feature extraction projection matrix ; Step 8: Calculate the T² statistic of the training set according to formula (25), and determine the control limit of the T² statistic according to formula (26) using a mixed threshold strategy. .

[0063] S302, Online Detection Process Step 1: Collect transition data from normal to faulty operation of the pneumatic control valve ; Step 2: Predict the flow coefficient based on the GPR model built in Step 3 during offline training. ; Step 3: Calculate the flow coefficient C of the test data according to formula (8). V residual ; Step 4: Flow coefficient C of test data V residual Control limits of the state detection index calculated in step 4 of offline training Compare the results to determine if a malfunction has occurred. Step 5: Calculate the row mean vector of the training kernel matrix using the offline training method from step 7. and global mean The kernel vectors of the test samples and training samples are calculated according to formula (27). And the kernel vector is centered according to formula (28). ; Step 6: Extract the projection matrix using the slow features calculated in step 7 of offline training. The slow features of the test samples are calculated according to formula (29). ; Step 7: Calculate the mean vector and covariance matrix of the test data according to formula (24), and calculate the T² statistic of the test data according to formula (25); Step 8: T² statistic of test data and control limits of state detection index calculated in step 8 of offline training Compare the results to determine if a malfunction has occurred. Step 9: Update the continuous counter according to formula (32). For fault detection of fast layer and slow layer, apply the continuous filtering mechanism according to formula (33). If the statistical index exceeds the control limit and the number of consecutive abnormalities reaches the set value, the control valve is considered to have failed, and a fault warning is issued; otherwise, the control valve is considered to be in normal operation and return to step 1.

[0064] This invention uses the pneumatic control valve benchmark simulation platform (Development and Application of Methods for Actuator Diagnosis in Industrial Control Systems, DAMADICS) as a verification platform to detect common faults in the operation of pneumatic control valves, verify the fault detection performance of this method for pneumatic control valves, and compare it with KPCA, KECA and MS-MWKECA multivariate statistical analysis methods.

[0065] This invention uses Fault Detection Rate (FDR) and False Alarm Rate (FAR) as evaluation metrics to assess the fault detection effectiveness of all methods. FDR is the proportion of correctly detected fault samples out of the total number of fault samples, and FAR is the proportion of normal samples that are incorrectly identified as fault samples. (34) In the formula, This indicates the number of samples that actually failed and were correctly detected. This represents the number of samples that actually had a fault but were not detected.

[0066] (35) In the formula, This indicates the number of normal samples that were misclassified as faulty. This indicates the number of samples that were correctly identified as normal samples.

[0067] Comparative Analysis of Detection Results. Table 2 summarizes the fault detection results of KPCA, KECA, MS-MWKECA, and this method for six different fault modes of pneumatic control valves in the DAMADICS platform. The best detection result for each fault mode is indicated in bold. The VoA statistic in the KECA method is an angle monitoring index designed by analyzing the structural characteristics between the data after KECA transformation. The data in the table show that the overall performance of the KPCA method is mediocre. Its T² statistic has a low fault detection rate and a large number of missed detections. Although its SPE statistic reaches 98.71% for F3, its detection rate for other faults is low, making it difficult to achieve stable monitoring. The VoA statistic of the KECA method shows significant fluctuations. Although it achieves a high detection rate of over 90% for F2 and F3 faults, the detection rate for faults F4, F5, and F6 is only around 61% to 64%, indicating that the method is more sensitive to specific fault types. The MS-MWKECA method is significantly improved compared to the previous two, with a detection rate of 95.36% for F2 faults, but its detection rate is only 69.36% when facing fault F5. In contrast, the CV residual statistics constructed by our method remain above 72% for all six types of faults, with a detection rate of over 98% for five types of faults, demonstrating extremely high sensitivity and robustness. At the same time, the T² statistics of our method also remain at a high level of 73% to 98%, and the two complement each other effectively, providing excellent and stable detection capabilities for different types of faults, thus verifying the effectiveness and advancement of the proposed method.

[0068] Table 2 Fault Detection Rate (FDR) of Different Methods Table 3 compares the average false alarm rates of the four methods. The CV residual statistic of this invention achieves a zero false alarm rate, demonstrating its excellent balance between fault sensitivity and stability under normal operating conditions. Meanwhile, the false alarm rate of the T² statistic of this method is 1.20%, which is within a reasonable range and effectively complements the CV residual. This invention outperforms the KPCA, KECA, and MS-MWKECA methods in fault detection rate, and also excels in false alarm rate control, achieving an optimal trade-off between high detection rate and low false alarm rate. This verifies the reliability and practicality of the proposed method in industrial process monitoring.

[0069] Table 3 Comparison of the average false alarm rate (FAR) of the four methods To thoroughly and intuitively evaluate the detection performance of this invention, the detection results are visualized and compared, where the red dashed line represents the control limit, the blue solid line represents the detection index, and the red dotted line represents the fault occurrence time point 900s.

[0070] The detection results of fault F1 by KPCA, KECA, MS-MWKECA, and this invention are as follows: Figures 4-7 As shown, combining the detection results in Table 2 and Figure 4 It can be seen that although the SPE and T² statistics of the KPCA method respond to abnormal operating conditions after the fault occurs, they both exhibit periodic oscillations, causing the statistics to frequently fall below the control limits in each cycle, resulting in extremely discontinuous detection signals and an inability to form stable fault alarms. Figure 5 The VoA statistic of the KECA method exhibits similar pulse-like fluctuations. Although its instantaneous peak value far exceeds the control limit, the instability of the signal causes it to cross the control limit multiple times, resulting in missed detections and making it difficult to maintain continuous and effective monitoring. Figure 6 Although the T² statistic of the MS-MWKECA method shows a jump in amplitude at the fault time, the overall signal strength is significantly insufficient, hovering around the control limits and exhibiting strong random noise. It lacks significant fault discrimination and struggles to achieve high-confidence detection of fault F1. In contrast, this invention… Figure 7 Both statistics exhibit ideal detection characteristics, C V The residuals respond rapidly after the fault occurs and continuously and stably exceed the control limits, with almost no missed detections. The T² statistic also shows a significant and continuous exceedance of the limit within the fault interval, achieving a detection rate of 97.73%. This method demonstrates stronger fault sensitivity and more stable detection performance in the detection of fault F1, fully verifying the effectiveness and superiority of the proposed method.

[0071] The detection results of four detection methods for the typical small gradual change fault F5 are as follows: Figures 8-11 As shown, Figure 8 Although the T² and SPE statistics of the KPCA method show a gradual trend of slowly increasing with the increase of fault intensity, they are affected by strong periodic noise signals and fail to effectively exceed the limit in the early stage of the fault, resulting in a low detection rate. Figure 9 The VoA statistic of the KECA method exhibits an extreme impulse response. Although the peak value at some moments far exceeds the control limit, the lack of signal continuity causes the statistic to frequently drop to the safe region between pulses, resulting in an actual detection rate of only 61.09%. Figure 10 The T² statistic of the MS-MWKECA method is insufficiently sensitive to small, gradual changes in signals. Although the curve rises overall after a fault occurs, it oscillates around the control limits, failing to extract features significantly different from normal operating conditions, resulting in a detection rate of only 69.36%. In contrast, Figure 11 The method exhibits unique advantages in detecting gradual faults, C V The residual statistic can keenly capture abnormal signals where the control variable deviates from the expected behavior. Several significant impulsive exceedance responses appear during the fault development process, achieving a high detection rate of 99.55%. The T² statistic shows a gradually increasing trend with fault evolution, progressing from small fluctuations in the early stages to large and sustained exceedances in the later stages, with a detection rate of 73.64%. Therefore, facing gradual faults like F5, which are difficult to detect, this invention, through deep modeling of process dynamics and control behavior, can effectively identify the gradual evolution of the fault, verifying its superior detection capability under complex operating conditions.

[0072] Experimental Platform and Data Acquisition. To verify the effectiveness of the proposed method in a real industrial environment, this invention utilizes an I-shaped pipe network experimental platform from a certain enterprise to conduct fault detection experiments. The I-shaped pipe network experimental platform is as follows: Figure 12 As shown, the pipeline system consists of two branch pipelines, an intermediate branch pipeline, and three pneumatic control valves, simulating the actual working conditions of fluid transportation and distribution in a typical process industry. Various high-precision sensors are deployed, including a MIK-PX400 pressure transmitter, a DKLD-DN100 intelligent electromagnetic flowmeter, and an AEP series intelligent valve positioner. The experiment selected pneumatic control valve 2 in the pipeline system as the research object. This control valve is an ATS model with a nominal diameter of 100mm, a rated stem stroke of 30mm, and an equal percentage flow characteristic. It is driven by a pneumatic diaphragm actuator and equipped with an intelligent valve positioner to achieve precise valve position control.

[0073] During the experiment, data acquisition cards were used to collect valve stem displacement signals, upstream pressure signals, downstream pressure signals, and flow signals of the control valve. Three typical operating conditions were designed and simulated: 1) Normal fixed-opening data, collecting different opening values ​​from 5% to 100% in 5% increments during normal valve operation; 2) Valve external leakage fault, simulating the fault by adjusting the opening of valves in intermediate pipeline branches to control the external leakage of the pneumatic control valve; 3) Positioner feedback fault, simulating the fault by adjusting the preload of the intelligent valve positioner feedback linkage to create a deviation between the positioner's feedback signal and the actual opening. Data at different opening values ​​under normal operating conditions were used to train a Gaussian process regression model, data at 100% normal operating condition opening were used to train a kernel-slow feature analysis model, and transitional data from 100% normal operating condition opening to the fault state were used to verify the algorithm's detection performance. This experimental platform realistically reflects the dynamic characteristics and nonlinear disturbances of process industry pipeline systems, effectively verifying the applicability and robustness of the proposed method in complex industrial environments.

[0074] The experimental detection results of the process fault detection method proposed in this invention for external leakage faults in pneumatic control valves are as follows: Figure 13 As shown, the red dashed line represents the control limit, and the blue solid line represents the detection index. The C of the fast layer... V The residuals exhibited extremely high sensitivity; as the leakage increased, the residual curve exceeded the lower control limit at approximately the 1880th sampling point, directly reflecting the deviation of the actual flow coefficient from the predicted trajectory of the GPR model. Simultaneously, the slow layer... After a transition period, the statistic exceeded the control limit near the 2200th sampling point and showed a continuous upward trend, effectively capturing the global characteristics of system degradation. This C... V The rapid response of residuals to sudden changes in process variables and The combination of statistics with system state offsets verifies the superior performance of the two-layer fusion mechanism in handling external leakage faults, combining rapid alarm and continuous monitoring.

[0075] The detection results for feedback faults in pneumatic control valve positioners are as follows: Figure 14 As shown, a fault occurs at approximately 2000 sampling points, in the C of the fast layer. V The residuals respond rapidly, immediately exceeding the control limit and exhibiting a step-like upward trend, accurately reflecting the nonlinear prediction error of the flow coefficient caused by loose mechanical connections. Meanwhile, the slow layer T... 2 The statistical results also triggered an alarm simultaneously, far exceeding the control limit, indicating that the algorithm effectively captured the continuous degradation characteristics caused by feedback loop mismatch within the system. Experimental results fully demonstrate that the proposed process fault detection method can not only accurately identify such signal deviation faults caused by mechanical coupling, but also maintain high detection stability during the fault's duration.

[0076] This invention addresses the problem of feature extraction and performance balance in fault detection of pneumatic control valves under complex operating conditions. It proposes and verifies a control valve process fault detection scheme based on mechanistic residual and kernel slow feature analysis. This is achieved by integrating C++ based on Gaussian process regression. V Residual analysis and activity-based KSFA features enable comprehensive monitoring of both transient and gradual degradation faults. Simulation results on the DAMADICS platform show that C VThe residual detection rate exceeded 72% in all six typical fault types, with five exceeding 98%, achieving a zero false alarm rate, verifying its excellent stability while ensuring high sensitivity. The T² statistic maintained a detection rate ranging from 73% to 98%, complementing the fast layer and enhancing the detection architecture's adaptability to different fault modes. Comparative experiments with KPCA, KECA, and MS-MWKECA multivariate statistical analysis methods confirmed that the proposed method has significant advantages in both detection rate and false alarm rate. Furthermore, the introduced dual threshold and continuous judgment mechanism effectively suppressed false alarms caused by transient noise interference, keeping the average false alarm rate below 2%, meeting the stringent reliability requirements of industrial monitoring systems. Finally, to verify the effectiveness of the proposed method in a real industrial environment, fault detection experiments were conducted using an I-shaped pipeline network experimental platform of a certain enterprise, simulating two typical fault types. The method was able to significantly distinguish between normal operation and fault states in the test data, further verifying its accurate identification capability and engineering application value for various fault states in real industrial scenarios.

[0077] This invention solves the problem of significant nonlinear coupling characteristics in the prior art. Due to the complex mapping between flow rate, pressure and opening degree, the simple model will produce large deviations under normal operating conditions, which will easily cause false alarms. The prior art also has the problem that industrial field measurement noise and process disturbances are unavoidable. The algorithm needs to have strong anti-interference ability while ensuring fault sensitivity. In addition, early degradation characteristics are easily masked by normal fluctuations. Traditional threshold methods often relax the restrictions in order to reduce false alarms, resulting in missed detections.

[0078] The main contributions of the method of this invention include: (1) utilizing the nonparametric Bayesianness and adaptive kernel function learning capability of Gaussian Process Regression (GPR) to realize the control valve flow coefficient C V (1) High-precision fitting; (2) By introducing the feature activity index to optimize the selection strategy of slow features, noise interference is suppressed while preserving the information of the gradual change trend of the system, and the ability of slow features to characterize performance degradation faults is improved. (3) A fast-slow dual-layer collaborative detection architecture is designed, and the fast layer utilizes C V The residual monitoring control performance is immediately abnormal. The slow layer uses the T² statistic extracted by KSFA to capture the gradual degradation of the system state. The two detection mechanisms complement each other effectively, enhancing the comprehensive detection capability for different fault modes. (4) A dual threshold and continuous judgment mechanism is introduced. By setting dual thresholds for early warning and alarm, and combining them with continuous over-limit logic, false alarms caused by transient noise are effectively avoided while ensuring high detection sensitivity.

[0079] Based on the same inventive concept, the present invention also provides a fault detection system for a pneumatic control valve, comprising: The data acquisition module is used to collect physical monitoring variables in real time during the operation of the target pneumatic control valve. These physical monitoring variables include valve stem displacement, valve inlet pressure, valve outlet pressure, and fluid volumetric flow rate.

[0080] The calculation module is used to calculate the actual observed flow coefficient of the target pneumatic control valve based on the physical monitoring variables and the flow coefficient formula; input the valve stem displacement into a pre-trained Gaussian process regression (GPR) model and output the predicted observed flow coefficient; calculate the error value between the actual observed flow coefficient and the predicted observed flow coefficient, compare the residual value with the preset fast layer control limit, generate a binary fast anomaly preliminary flag, and generate a fast layer anomaly alarm signal based on the fast anomaly preliminary flag; calculate the standard deviation of the physical monitoring variables according to a fixed sliding window, and concatenate the physical monitoring variables at each time step with the standard deviation at the corresponding time step to form a multidimensional enhanced feature vector; input the multidimensional enhanced feature vector into a pre-trained KSFA model and output the Hotelling T² statistic, and generate a slow layer anomaly alarm signal based on the T² statistic.

[0081] The detection module is used to detect faults in the target pneumatic control valve based on the high-speed layer abnormal alarm signal and the slow-speed layer abnormal alarm signal.

[0082] This invention also provides a computer device. At the hardware level, the computer device includes a processor, an internal bus, a network interface, memory, and non-volatile memory, and may also include other hardware required for business operations. The processor reads the corresponding computer program from the non-volatile memory into the memory and then runs it to implement the aforementioned fault detection method for the pneumatic control valve.

[0083] The present invention also provides a computer-readable storage medium storing a computer program that can be used to execute the above-described fault detection method for pneumatic control valves.

[0084] Specific limitations of the computational system for the fault detection method of pneumatic control valves can be found in the limitations of the fault detection method for pneumatic control valves mentioned above, and will not be repeated here. Each module in the aforementioned fault detection system for pneumatic control valves can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in the processor of a computer device in hardware form or independent of the processor, or stored in the memory of a computer device in software form, so that the processor can call and execute the corresponding operations of each module.

[0085] The technical features of the above embodiments can be combined arbitrarily. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as the combination of these technical features does not contradict each other, it should be considered within the scope of this specification. Furthermore, the above embodiments only illustrate several implementation methods of this application, and their descriptions are relatively specific and detailed, but they should not be construed as limiting the scope of the invention patent. It should be noted that those skilled in the art can make several modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this patent application should be determined by the appended claims.

Claims

1. A method for fault detection of a pneumatic control valve, characterized in that, include: The physical monitoring variables of the target pneumatic control valve under real-time operating conditions are collected. The physical monitoring variables include valve stem displacement, valve inlet pressure, valve outlet pressure, and fluid volume flow rate. Based on the physical monitoring variables, the actual observed flow coefficient of the target pneumatic control valve is calculated according to the flow coefficient formula; the valve stem displacement is input into the pre-trained Gaussian process regression (GPR) model, and the predicted observed flow coefficient is output; the error value between the actual observed flow coefficient and the predicted observed flow coefficient is calculated, the residual value is compared with the preset fast layer control limit, a binary fast anomaly preliminary flag is generated, and a fast layer anomaly alarm signal is generated based on the fast anomaly preliminary flag. The standard deviation of the physical monitoring variables is calculated using a fixed sliding window. The physical monitoring variables at each time point are concatenated with the standard deviations at the corresponding time points to form a multidimensional enhanced feature vector. The multidimensional enhanced feature vector is input into a pre-trained KSFA model, which outputs the Hotelling T² statistic. An abnormal alarm signal for the slow layer is generated based on the T² statistic. Based on the high-speed layer abnormality alarm signal and the slow-speed layer abnormality alarm signal, the fault detection of the target pneumatic control valve is realized.

2. The fault detection method for a pneumatic control valve according to claim 1, characterized in that, The step of generating a rapid layer anomaly alarm signal based on a rapid anomaly preliminary flag specifically includes: generating a rapid layer anomaly alarm signal when the rapid anomaly preliminary flag continuously reaches or exceeds a first preset threshold; the step of generating a slow layer anomaly alarm signal based on T² statistics specifically includes: setting the preliminary anomaly flag to 1 if the T² statistics exceed a preset slow layer control limit; and generating a slow layer anomaly alarm signal when the anomaly flag continuously reaches or exceeds a second preset threshold.

3. The fault detection method for a pneumatic control valve according to claim 1, characterized in that, The Gaussian process regression (GPR) model is trained using a Gaussian process regression algorithm with a composite kernel function. The kernel function parameters are optimized by maximizing the marginal likelihood function to obtain the GPR model.

4. The fault detection method for a pneumatic control valve according to claim 1, characterized in that, During training, the KSFA model calculates the RBF kernel matrix of the multidimensional enhanced feature vectors of the training samples, centers the RBF kernel matrix, and calculates the time difference kernel matrix based on the centered RBF kernel matrix. Calculate the generalized eigenvalues ​​of the time difference kernel matrix and the centered RBF kernel matrix; select the sample vectors corresponding to the k smallest generalized eigenvalues ​​to form a slow feature extraction projection matrix; obtain the Hotelling T² statistic based on the slow feature extraction projection matrix to complete the KSFA model training.

5. The fault detection method for a pneumatic control valve according to claim 1, characterized in that, Based on the fast-layer abnormal alarm signal and the slow-layer abnormal alarm signal, the fault detection of the target pneumatic control valve is realized. Specifically, the fast-layer abnormal alarm signal is used to detect transient sudden faults, and the slow-layer abnormal alarm signal is used to detect gradual degradation faults. If any alarm signal is triggered, it is determined that there is a fault in the target pneumatic control valve.

6. The fault detection method for a pneumatic control valve according to claim 1, characterized in that, The control limit for the fast layer is set based on the 3σ principle, with the first preset threshold being 5. The control limit for the slow layer is determined using a hybrid threshold strategy, set by the mean and standard deviation of the T² statistic of the training set, with the second preset threshold being 15.

7. The method for fault detection of a pneumatic control valve according to claim 1, characterized in that, The step of calculating the standard deviation of the physical monitoring variables using a fixed sliding window and concatenating the physical monitoring variables at each time point with the standard deviations at the corresponding time point into a multi-dimensional enhanced feature vector specifically includes: obtaining the valve pressure difference between the valve inlet pressure and the valve outlet pressure; calculating the standard deviations of the valve stem displacement, valve pressure difference, and fluid volumetric flow rate respectively; and concatenating the valve stem displacement, valve pressure difference, and fluid volumetric flow rate at each time point with the standard deviations at the corresponding time point into a 6-dimensional enhanced feature vector.

8. A fault detection system for a pneumatic control valve, characterized in that, include: The data acquisition module is used to collect physical monitoring variables of the target pneumatic control valve in real time during operation. The physical monitoring variables include valve stem displacement, valve inlet pressure, valve outlet pressure, and fluid volume flow rate. The calculation module is used to calculate the actual observed flow coefficient of the target pneumatic control valve based on the physical monitoring variables and the flow coefficient formula; input the valve stem displacement into a pre-trained Gaussian process regression (GPR) model and output the predicted observed flow coefficient; calculate the error value between the actual observed flow coefficient and the predicted observed flow coefficient, compare the residual value with the preset fast layer control limit, generate a binary fast anomaly preliminary flag, and generate a fast layer anomaly alarm signal based on the fast anomaly preliminary flag; calculate the standard deviation of the physical monitoring variables according to a fixed sliding window, and concatenate the physical monitoring variables at each time step with the standard deviation at the corresponding time step to form a multidimensional enhanced feature vector; input the multidimensional enhanced feature vector into a pre-trained KSFA model and output the Hotelling T² statistic, and generate a slow layer anomaly alarm signal based on the T² statistic; The detection module is used to detect faults in the target pneumatic control valve based on the high-speed layer abnormal alarm signal and the slow-speed layer abnormal alarm signal.

9. A computer device, comprising a memory, a processor, and a computer program stored in the memory, characterized in that, The processor executes the computer program to implement the steps of the method according to any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is loaded by the processor, it is able to perform the steps of the method according to any one of claims 1 to 7.

Images