A real-time online monitoring method for a pneumatic control valve

Abstract

The present invention relates to the technical field of control valves, and specifically to a real-time online monitoring method for a pneumatic control valve. The present invention first collects the operation data of the pneumatic control valve, transmits it to the edge computing node for preprocessing and lightweight processing, and then generates spectral data. Next, the characteristic data in the spectral data is extracted and compressed to form the control valve characteristic data. At the central node, multi-source data fusion and dimensionality reduction are performed on the control valve characteristic data to generate the principal component fusion characteristics. Then, the key characteristics in the fusion characteristics are extracted to generate a fault characteristic set, and a random forest model is used to obtain the fault identification result. An adaptive model update mechanism is used to adjust and update the random forest model; finally, based on the historical operation data of the pneumatic control valve, the random forest model is extended and trained to improve the multi-fault identification ability. The present invention combines edge lightweight data processing with deep learning to improve the accuracy, robustness, and real-time performance of the real-time monitoring of the pneumatic control valve.

Description

Technical Field

[0001] The invention relates to the technical field of online monitoring of pneumatic control valves, and in particular to a real-time online monitoring method of pneumatic control valves. Background Art

[0002] Pneumatic control valve is an automatic control device that is driven by compressed air to adjust the flow, pressure, temperature or liquid level of the fluid. Its principle is to receive the pneumatic signal of the control system through the pneumatic actuator, drive the valve core to move or rotate in the valve body, change the valve opening, and thus control the flow of the fluid. Pneumatic control valves are widely used in the petroleum, chemical power, food, and pharmaceutical industries for precise regulation in production process control due to their advantages such as fast response, high precision, and good adaptability to harsh environments.

[0003] Real-time monitoring of pneumatic control valves can detect potential faults in a timely manner, avoid unexpected accidents, reduce downtime and reduce maintenance costs. At the same time, real-time monitoring helps to accurately control the operating status of the valve and extend the life of the pneumatic control valve equipment. The existing technology mainly collects key operating parameters through sensors, and uses signal processing technology and machine learning methods to analyze the data to achieve real-time monitoring of pneumatic control valves. This method can continuously monitor the operating status of pneumatic control valves and achieve high fault diagnosis accuracy and low missed diagnosis rate, but there are still some shortcomings.

[0004] Existing technologies usually detect and identify faults for specific types of control valves or single fault modes by integrating the collected data into a central processing terminal for processing and analysis. In actual pneumatic control valve application scenarios, multiple types of control valves are usually used in combination and multiple types of faults occur simultaneously. Under such complex working conditions, the data calculation complexity of existing technologies during real-time monitoring increases, and the processing difficulty increases. These problems result in the accuracy, robustness and real-time performance of real-time monitoring of pneumatic control valves still need to be improved.

[0005] Therefore, a real-time online monitoring method for pneumatic control valves is proposed. Summary of the invention

[0006] The object of the present invention is to provide a real-time online monitoring method for a pneumatic control valve. First, the operation data of the pneumatic control valve is collected and transmitted to an edge computing node for preprocessing and lightweight processing to generate spectrum data. Then, feature calculation is performed on the spectrum data to extract and compress the feature data. At the central node, multi-source data fusion and dimensionality reduction are performed on the feature data to generate the principal component fusion feature. Next, key features are extracted from the fusion feature to generate a fault feature set for obtaining a fault recognition model, and an adaptive model update mechanism is implemented to adjust and update the fault recognition model. Finally, based on the historical operation data of the pneumatic control valve and combined with deep learning methods, the fault recognition model is trained and optimized to improve the multi-fault recognition ability. The present invention improves the accuracy, robustness, and real-time performance of the real-time monitoring of the pneumatic control valve through edge lightweight data processing combined with deep learning.

[0007] To achieve the above object, the present invention provides a real-time online monitoring method for a pneumatic control valve, including:

[0008] Collect the real-time signal data during the operation of the pneumatic control valve and transmit it to the edge computing node for processing; the real-time signal data includes air chamber pressure data, valve stem displacement data, temperature data, and vibration data;

[0009] Perform preprocessing and lightweight processing on the real-time signal data to generate the control valve spectrum data;

[0010] Extract the key frequency components in the control valve spectrum data through feature calculation to generate the control valve feature data and transmit it to the central computing node;

[0011] Perform principal component extraction and dimensionality reduction on the control valve feature data to obtain the principal component feature;

[0012] Perform multi-source data fusion on the principal component feature to generate the principal component fusion feature;

[0013] Analyze and extract key features from the fusion feature to generate the control valve fault feature set;

[0014] Use a random forest model to classify and identify the control valve fault feature set to obtain the fault recognition result;

[0015] Use an adaptive model update mechanism to adjust and update the random forest model;

[0016] Collect the historical operation data of the pneumatic control valve and the fault recognition result, and perform extended training and optimization on the random forest model to obtain a multi-fault recognition model.

[0017] Further, the preprocessing includes low-pass filtering, baseline correction, and outlier filtering;

[0018] x final x(t) = F outlier (F base (F filt (x(t)));

[0019] Wherein, x(t) is the signal data at time t, and x final (t) is the data signal after preprocessing, and F filt (·) is a low-pass filtering operation, and F base (·) is a baseline correction operation, and F Outlier (·) is an outlier detection and filtering operation.

[0020] Furthermore, the generation process of the regulating valve spectrum data is completed by Fourier transform;

[0021] The preprocessed data signal is converted into a frequency-domain signal through fast Fourier transform to generate the regulating valve spectrum data; the conversion formula of the frequency-domain signal is:

[0022]

[0023] Wherein, X(f) is the regulating valve spectrum data, j is the imaginary unit, f is the frequency component, and x final (i) is the data signal after preprocessing, and N is the total number of data points.

[0024] Furthermore, feature calculations are performed on the regulating valve spectrum data, including feature extraction, real-time analysis, and compression;

[0025] Through f max = argmax f |X(f)|, the maximum frequency f corresponding to the highest amplitude is extracted from the regulating valve spectrum data max ; the power spectral density PSD(f) of the regulating valve spectrum data is calculated by PSD(f) = |X(f)| 2 / N; the root mean square value RMS of the regulating valve spectrum data is calculated through ;

[0026] Abnormal monitoring is performed on the maximum frequency f max and the root mean square value RMS. When the detected feature value exceeds the normal range, it is marked as abnormal;

[0027] The maximum frequency, root mean square value, and power spectral density are compressed using the function X edge = compress(f max , PSD(f), RMS) to obtain the regulating valve feature data; the regulating valve feature data is transmitted to the central computing node for processing.

[0028] Furthermore, the steps of the multi-source data fusion include principal component analysis and weighted fusion;

[0029] Dimensionality reduction is performed on the regulating valve characteristic data from different edge computing nodes using the principal component analysis method to generate principal component features;

[0030] Z = A·[P edge (f), d edge (f), T edge (f), a edge (f)]

[0031] where Z is the principal component feature, A is the feature vector matrix of the principal component analysis, and P edge (f), d edge (f), T edge (f) and a edge (f) are the characteristic data of pressure, displacement, temperature, and vibration, respectively;

[0032] The principal component features are weighted and fused according to the weights to generate principal component fusion features;

[0033] F = ω P ·Z P + ω d ·Z d + ω T ·Z T + ω a ·Z a ;

[0034] where Z P 、Z d 、Z T and Z a are the principal component features of pressure, displacement, temperature, and vibration, respectively, and ω P 、ω d 、ω T and ω a are the weights corresponding to the principal component features, respectively.

[0035] Furthermore, the steps for obtaining the fault feature set are as follows:

[0036] Analyze the fusion features, and extract the key features related to the fault through F fault = E(F) to generate the fault feature set F fault , where E(·) is the feature extraction function.

[0037] Furthermore, the steps for obtaining the fault recognition model include;

[0038] Adopt the random forest C fault = RF(F fault) Classify the fault feature set and output the fault recognition result; C fault is the fault recognition result, and RF(·) is the random forest model.

[0039] Further, the steps of the adaptive update mechanism are as follows;

[0040] Calculate the difference between the fault recognition result and the true fault type to determine the trigger condition of the adaptive mechanism;

[0041] Collect the latest operation data of the pneumatic control valve to generate a real-time fault feature set;

[0042] Update the parameters of the random forest model according to the real-time fault feature set by using an adaptive online learning algorithm.

[0043] θ RF,new = θ RF,ori - η RF · Δ θRF Φ(F new, , C fault );

[0044] Among them, θ RF,ori and θ RF,new are the parameters of the random forest model before and after update respectively; η RF is the learning rate, Φ(F new, , C fault ) is the loss function, Δ θRF is the change rate of the loss function, F new, is the real-time fault feature set.

[0045] Further, the training process of the multi-fault recognition model is as follows:

[0046] Collect the historical data of the pneumatic control valve, including historical feature data and the corresponding fault recognition results.

[0047] Conduct extended training on the random forest model to extract the correlation features between the historical feature data and the fault recognition results;

[0048]

[0049] Among them, DNN multi is the correlation feature, FC(·) is the three-layer convolution operation, x' history (m) is the m-th historical feature, and C history (p) is the fault recognition result of the p-th one.

[0050] The extended training process includes:

[0051] Input the historical feature data and the corresponding fault identification results into a three-layer convolutional layer to extract fault-related features and output the first associated feature map;

[0052] Use max pooling to reduce the dimension of the fault-related features, reduce the size of the first associated feature map and retain important information, and output the second associated feature map;

[0053] Unroll the second associated feature map into a vector and extract the global fault-related features;

[0054] Through the softmax activation function, convert the global fault-related features into a probability distribution of fault types to obtain the associated features;

[0055] Adjust the random forest model through the associated features to obtain a multi-fault identification model;

[0056] Furthermore, the optimization process of the multi-fault identification model is as follows:

[0057] Collect the latest operation data of the pneumatic control valve and obtain the multi-fault identification results using the multi-fault identification model;

[0058] Calculate the difference between the multi-fault identification results and the true multi-fault types in real time, and optimize and adjust the multi-fault identification model according to the difference;

[0059]

[0060] Among them, is the parameter of the preliminary multi-fault identification model, η RF is the learning rate, Φ(F train ,C multi ) is the loss function, is the change rate of the loss function, F train is the fault feature set of the historical operation data, C multi is the true multi-fault type corresponding to the historical operation data.

[0061] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0062] 1. The present invention first preprocesses the operation data of the pneumatic control valve, converts the preprocessed time-domain signal into a frequency-domain signal through fast Fourier transform, generates the spectrum data of the control valve, and extracts the key frequency components therefrom. Through this step, the complex time-domain signal is converted into a more easily analyzable frequency-domain signal, and the key frequency information reflecting the operation state of the control valve is extracted, solving the problem of high computational complexity in the analysis of time-domain signals in the prior art, and providing data support for improving the real-time performance and accuracy of online monitoring of pneumatic control valves.

[0063] 2. The present invention transfers the data feature extraction task to the edge computing node. By extracting key feature data and performing real-time analysis and monitoring, it detects abnormal features and marks outliers. After the key feature data is compressed and dimension-reduced, the principal component fusion features are generated through multi-source data fusion. This method solves the computational compliance problem in the prior art when centrally processing data and improves the data processing efficiency. At the same time, multi-source data fusion simplifies the data structure, enhances the processing of multi-source information and the fault detection ability, and improves the robustness and real-time performance of the online monitoring of pneumatic control valves.

[0064] 3. The present invention generates a fault feature set by extracting fault-related key features from the principal component fusion features; then uses a random forest model to classify and identify the fault feature set to obtain the fault identification result; then adjusts the model parameters in real time through an adaptive mechanism, and based on historical operation data and the fault identification result, performs extended training on the optimized random forest model, and finally adjusts and optimizes through the cross-entropy loss function to obtain a multi-fault identification model. This method solves the problem in the prior art that it is difficult to identify concurrent faults by combining fault feature extraction and random forest classification. By comprehensively applying adaptive update and deep learning training, the accuracy of the real-time monitoring of pneumatic control valves is comprehensively improved. BRIEF DESCRIPTION OF THE DRAWINGS

[0065] Figure 1 It is a schematic flowchart of a real-time online monitoring method for a pneumatic control valve provided by an embodiment of the present invention;

[0066] Figure 2 It is a schematic flowchart of the acquisition process of the control valve feature data provided by an embodiment of the present invention;

[0067] Figure 3 It is a schematic flowchart of the training and optimization process of the multi-fault identification model provided by an embodiment of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0068] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

[0069] The present invention provides a real-time online monitoring method for a pneumatic control valve. For the specific method flowchart, refer to Figure 1 .

[0070] Embodiment 1

[0071] As an implementation manner of the present invention, refer to Figure 1S10 in it is used to collect real-time data during the operation of the pneumatic control valve, including air chamber pressure data, valve stem displacement data, temperature data, and vibration data, and transmit them to the edge node for preprocessing and lightweight preliminary processing to generate the control valve spectrum data;

[0072] The edge nodes are deployed in the valve assembly area of each pneumatic control valve, and each edge node is configured with edge computing equipment; the operation data of the pneumatic control valve is collected by sensors and transmitted to the edge computing equipment through wired devices;

[0073] The preprocessing includes low-pass filtering, baseline correction, and outlier detection and filtering; for the original data signal, first, the high-frequency noise is removed through a low-pass filter;

[0074]

[0075] x filtered (t) is the filtered signal, h k is the filter coefficient, N is the filter order, and x(t) is the original data signal of the pneumatic control valve.

[0076] The baseline drift in the data signal is eliminated by the moving average method, and the corrected signal is:

[0077]

[0078] x corrected (t) is the signal after baseline correction, and M is the window size of the moving average.

[0079] The 3-sigma method is used to identify and exclude outliers;

[0080]

[0081] x final (t) is the time-domain signal after preprocessing, μ x is the average value of the signal after baseline correction, σ x is the standard deviation of the signal after baseline correction, and NA indicates that the outliers are filtered.

[0082] In this embodiment, through low-pass filtering, baseline correction, and outlier detection and filtering, the influence of noise, drift, and abnormal data caused by non-faults is reduced in the preprocessing step, ensuring the stability and reliability of the data, and providing high-quality input for subsequent analysis and processing.

[0083] The generation of the control valve spectrum data mainly uses the fast Fourier transform method;

[0084] For the preprocessed time-domain signal, through fast Fourier transform, it is converted into a frequency-domain signal to generate the spectrum data of the regulating valve; the conversion formula of the frequency-domain signal is:

[0085]

[0086] where X(f) is the spectrum data of the regulating valve, j is the imaginary unit, f is the frequency component, and x final (i) is the preprocessed data signal, and N is the total number of data points.

[0087] In this embodiment, the fast Fourier transform is used to convert the time-domain signal into a frequency-domain signal, making the complex time-domain data easier to analyze. In scenarios with obvious frequency characteristics, it not only improves the efficiency of data processing but also reduces the redundant analysis of the original data, providing support for improving the fast response ability of real-time monitoring of pneumatic regulating valves.

[0088] Further, referring to Figure 1 S20 in, S20 is used to perform feature calculation on the spectrum data of the regulating valve, further extract data features, and generate the feature data of the regulating valve;

[0089] Referring to Figure 2 , the feature calculation process includes feature extraction, real-time analysis, and compression;

[0090] Through f max = argmax f |X(f)| to extract the maximum frequency f corresponding to the highest amplitude from the spectrum data of the regulating valve max ;

[0091] Calculate the power spectral density PSD(f) of the spectrum data of the regulating valve through PSD(f) = |X(f)| 2 / N;

[0092] Through calculate the root mean square value RMS of the spectrum data of the regulating valve;

[0093] Perform anomaly monitoring on the maximum frequency f max and the root mean square value RMS. When the detected feature value exceeds the normal range, it is marked as abnormal;

[0094] Compress the maximum frequency, root mean square value, and power spectral density using the function X edge = compress(f max , PSD(f), RMS) to obtain the feature data of the regulating valve;

[0095] The feature data of the regulating valve is transmitted to the central computing node through a wireless device.

[0096] In this embodiment, by calculating the characteristics of the spectrum data of the regulating valve, including extracting the maximum power, power spectral density, and root mean square value, and performing real-time anomaly monitoring and data compression, the accuracy and real-time performance of data analysis are improved. This method can timely capture anomaly characteristics to ensure the rapid identification of potential problems of the pneumatic regulating valve; at the same time, by compressing key feature data, the storage and transmission burdens are reduced, and the data complexity of subsequent analysis is also reduced.

[0097] Further, referring to Figure 1 S30 in, S30 is used to perform principal component extraction and dimensionality reduction on the regulating valve feature data from the edge computing node, and then perform multi-source data fusion to generate principal component fusion features, including principal component analysis and weighted fusion;

[0098] Dimensionality reduction is performed on the regulating valve feature data from the edge computing node using the principal component analysis method to generate principal component features;

[0099] Z = A · [P edge (f), d edge (f), T edge (f), a edge (f)]

[0100] where Z is the principal component feature, A is the feature vector matrix of principal component analysis, and P edge (f), d edge (f), T edge (f) and a edge (f) are the feature data of chamber pressure, valve stem displacement, temperature, and vibration respectively;

[0101] The principal component features are weighted and fused according to the weights to generate principal component fusion features;

[0102] F = ω P ·Z P + ω d ·Z d + ω T ·Z T + ω a ·Z a ;

[0103] where Z P , Z d , Z T and Z a are the principal component features of pressure, displacement, temperature, and vibration respectively, and ω P , ω d , ω T and ω a are the weights corresponding to the principal component features respectively. The weights are set by professionals according to the actual application scenario.

[0104] In this embodiment, principal component analysis is performed on multi-source data signals such as pressure, displacement, temperature, and vibration to extract the most important principal component features. Then, different signals are weighted and fused, simplifying the data dimension, improving the efficiency and accuracy of signal processing, and retaining key fault feature information, which helps to improve the accuracy of fault identification.

[0105] Further, referring to Figure 1 S40 in

[0106] which is used to analyze the principal component fusion features and extract key features to generate a regulating valve fault feature set;

[0107] The steps for obtaining the fault feature set are as follows:

[0108] According to the fault modes of the pneumatic regulating valve, the fusion features are analyzed, and key features related to faults are extracted through F fault =E(F);

[0109] The fault types and key data features are listed in Table 1;

[0110] Table 1 Fault Types and Related Data Features

[0111]

[0112] Vibration fault feature: Detect abnormal frequency fluctuations or high-frequency components in the vibration signal;

[0113] Pressure fault feature: Analyze the maximum and minimum values in the time domain of the air chamber pressure signal to identify abnormal air chamber pressure fluctuations;

[0114] Temperature fault feature: Analyze the time-domain features of the temperature signal to determine whether the temperature exceeds the safe range;

[0115] Displacement fault feature: Monitor the displacement signal fluctuations through the time-domain and frequency-domain features of the displacement signal;

[0116] Signal change rate: Monitor sharp changes in the signal.

[0117] Generate a fault feature set F fault =[f max ,f min ,f mean ,f rms ,f freq ,f psd ,f rate , where f max and f min are the peak and minimum values of various signals respectively, and fmean and f rms are the mean and root mean square value of various signals respectively, reflecting the overall trend or steady state of the signals; f freq is the main frequency component used to identify periodic faults; f psd is the signal power spectral density, reflecting the energy distribution of the signal at different frequencies; f rate is the signal change rate used to identify signal mutations.

[0118] In this embodiment, the time-domain and frequency-domain features of multi-source data signals such as pressure, displacement, temperature, and vibration are fused to generate a fault feature set of the pneumatic control valve, covering all key operating parameters of the pneumatic control valve, providing rich features for the establishment of the subsequent fault identification model. By extracting fault-related features, this method removes redundant or useless data, simplifies the data analysis process, and improves the efficiency of subsequent fault identification.

[0119] The steps for obtaining the fault identification model include;

[0120] Construct a random forest model according to historical experience, and input the fault feature set F fault into the random forest model; each decision tree in the random forest is trained on different subsets of the fault feature set to form independent classification rules for identifying different fault modes;

[0121] Each decision tree in the random forest independently classifies and predicts the fault feature set, and each tree predicts the fault type corresponding to the current feature according to its classification rule;

[0122]

[0123] where T i represents the i-th decision tree, is the classification result of the i-th decision tree for the fault type.

[0124] The random forest uses a voting mechanism for the final judgment of the fault type to generate a fault identification result.

[0125]

[0126] Among them, mode is the majority vote mechanism of the random forest.

[0127] In this embodiment, a random forest model is used to integrate multiple decision trees for fault classification and identification, and the majority vote mechanism is adopted to improve the accuracy and robustness of fault identification. This method can handle complex multi-dimensional features, quickly classify and accurately identify fault features, ensuring real-time and high efficiency while improving the accuracy of fault detection.

[0128] Furthermore, referring to Figure 1S50 in it is used to adjust and update the random forest model by adopting an adaptive model update mechanism;

[0129] The steps of the adaptive update mechanism are as follows: regularly collect the latest operation data of the pneumatic control valve, and generate a real-time fault feature set F according to known feature extraction rules new ;

[0130] Use an adaptive online learning algorithm to update the parameters of the fault recognition model according to the real-time fault feature set;

[0131] Real-time adjust the parameters of the fault recognition model and update the model through online learning;

[0132]

[0133] Among them, θ RF,ori and θ RF,new are the parameters of the random forest model before and after update respectively; η RF is the learning rate, which is used to control the step size of model parameter adjustment; Φ(F new, , C fault ) is the loss function; is the gradient of the loss function with respect to the model parameters, indicating the direction of parameter adjustment; F new is the real-time fault feature set.

[0134] The loss function Φ(·) reflects the model classification error. Calculate the loss function according to the difference between the real-time data and the recognition result;

[0135]

[0136] Among them, C q is the label of the actual fault type, and c q is the fault type recognized by the fault recognition model;

[0137] When the value of Φ(F new , C fault-realtime ) increases, the adaptive mechanism will be automatically triggered, and the fault recognition model will be adjusted by introducing new feature data.

[0138] In this embodiment, the fault recognition model is dynamically adjusted through the adaptive update mechanism, so that the model can be continuously optimized, and the accuracy and robustness of fault type classification and recognition are improved. This method can adapt to the changes in the operating environment of the pneumatic control valve, reduce false alarms and missed alarms of faults, and ensure that the model maintains high-efficiency fault detection ability during long-term operation.

[0139] Furthermore, referring to Figure 1S60 in it is used to collect historical data of the pneumatic control valve, and the random forest model is extended and trained and optimized to obtain a multi-fault identification model.

[0140] Refer to Figure 3 , the training process of the multi-fault identification model is as follows:

[0141] Collect historical operation data D history =(X history , C history ), X history is the set of historical feature data of each signal, and C history is the set of the fault identification results;

[0142] Perform extended training on the random forest model to extract the correlation features between the historical feature data and the fault identification results;

[0143]

[0144] Among them, DNN multi is the correlation feature, FC(·) is a three-layer convolution operation, x' history (m) is the m-th historical feature, and C history (p) is the fault identification result of the p-th one.

[0145] The specific training process is as follows:

[0146] Input the historical feature data and the corresponding fault identification results into a three-layer convolutional layer, and use Z conv =σ(W conv *X history +b conv ) to extract the fault correlation features and output the first correlation feature map; Z conv is the first correlation feature map, W conv is the convolutional kernel weight, b conv is the bias, * represents the convolution operation, and σ(·) is the activation function

[0147] Perform dimensionality reduction on the first correlation feature map Z conv using max pooling, reduce the size of the first correlation feature map Z conv and retain the important information, and output the second correlation feature map;

[0148] Unfold the second correlation feature map into a vector, and pass it through Z fc =σ(W fc ·Z pool +b fc ) to extract the global fault correlation features; among them, Z fc is the global fault correlation feature, W fc is the weight, Zpool is the second associated feature map, b fc is the offset.

[0149] Through the softmax activation function, convert the fault-associated global feature into a probability distribution of fault types to obtain the associated feature;

[0150] Adjust the random forest model through the associated feature to obtain a multi-fault identification model;

[0151] In this embodiment, the fault identification model is extended and trained through a deep neural network, which can improve the accuracy of fault classification and the ability to identify multiple faults. The deep neural network is used to extract complex features from multi-dimensional signals, and the model can effectively identify concurrent faults to ensure efficient operation under complex working conditions.

[0152] The optimization process of the multi-fault identification model is as follows:

[0153] Collect the latest operation data of the pneumatic control valve, and use the multi-fault identification model to obtain the multi-fault identification result;

[0154] Use the cross-entropy loss function to calculate the difference between the multi-fault identification result and the true multi-fault type in real time, and optimize and adjust the multi-fault identification model according to the difference;

[0155]

[0156] Among them, is the parameter of the multi-fault identification model, η RF is the learning rate, Φ(F train , C multi ) is the loss function, is the change rate of the loss function, F train is the fault feature set of historical operation data, C multi is the true multi-fault type corresponding to the historical operation data.

[0157] In this embodiment, by optimizing the loss function and updating the model parameters, the fault detection accuracy and robustness of the multi-fault identification model are improved. Through optimization, the model can adapt to the complex working conditions of concurrent faults, improve the real-time performance and detection efficiency of the pneumatic control valve fault detection, and ensure rapid response and efficient processing of multiple faults.

[0158] In this embodiment, by combining edge computing with various data processing means, the fault detection and identification capabilities of pneumatic control valves are improved. Key features are extracted from the multi-dimensional signals of the operating data of pneumatic control valves to accurately identify various fault types. The adaptive mechanism ensures that the model is continuously optimized during operation, enhancing the robustness and accuracy of fault detection and reducing false alarms and missed detections. Edge computing and lightweight processing reduce the time cost and storage cost of data processing, improving the real-time performance and efficiency of real-time monitoring of pneumatic control valves and ensuring timely response and fault handling. Through the training and optimization of the multi-fault identification model, the operating system of pneumatic control valves has high adaptability and can still operate stably when concurrent faults occur in complex situations.

[0159] Embodiment 2

[0160] As an implementation manner of the present invention, this embodiment describes the specific implementation manner for the operating system of the pneumatic control valve in Chemical Plant A.

[0161] Collect the real-time data during the operation of the pneumatic control valve in Chemical Plant A, including air chamber pressure data, valve stem displacement data, temperature data, and vibration data, and collect 12 time points for each type of data;

[0162] Transmit the sensors to the edge node for preprocessing and lightweight preliminary processing to generate the spectrum data of the control valve;

[0163] The edge nodes are deployed in the valve assembly area of each pneumatic control valve, and each edge node is configured with edge computing equipment; the operating data of the pneumatic control valve is collected by sensors and transmitted to the edge computing equipment through wired devices;

[0164] Part of the collected data is listed in Table 2;

[0165] Table 2 Example of data collection part of Chemical Plant A

[0166]

[0167] For the original data signal, first remove the high-frequency noise through a low-pass filter; eliminate the baseline drift in the data signal by the moving average method, and identify and exclude outliers using the 3-fold standard deviation method;

[0168] For the preprocessed time-domain signal, through the fast Fourier transform, convert it into a frequency-domain signal to generate the spectrum data of the control valve;

[0169] Perform feature calculation on the spectrum data of the control valve to further extract data features and generate the feature data of the control valve; the feature calculation process includes feature extraction, real-time analysis, and compression;

[0170] Extract the maximum frequency corresponding to the highest amplitude from the spectrum data of the regulating valve; calculate the power spectral density of the spectrum data of the regulating valve; calculate the root mean square value of the spectrum data of the regulating valve;

[0171] Perform anomaly monitoring on the maximum frequency and the root mean square value, and mark it as abnormal when the detected characteristic value exceeds the normal range;

[0172] Compress the maximum frequency, the root mean square value, and the power spectral density using a function to obtain the characteristic data of the regulating valve, and transmit it to the central computing node.

[0173] Reduce the dimension of the characteristic data of the regulating valve from the edge computing node using the principal component analysis method to generate principal component features;

[0174] Perform weighted fusion on the principal component features according to the weights to generate principal component fusion features;

[0175] Analyze the fusion features and extract key features to generate a fault feature set of the regulating valve; use the random forest model to classify and identify the fault feature set of the regulating valve to obtain a fault identification model;

[0176] The steps for obtaining the fault feature set are as follows: According to the known fault modes of the pneumatic regulating valve, analyze the fusion features, and generate a fault feature set by extracting key features related to the faults

[0177] Vibration fault feature: Detect abnormal frequency fluctuations or high-frequency components in the vibration signal;

[0178] Pressure fault feature: Analyze the maximum and minimum values in the time domain of the air chamber pressure signal to identify abnormal fluctuations in the air chamber pressure;

[0179] Temperature fault feature: Analyze the time-domain characteristics of the temperature signal to determine whether the temperature exceeds the safe range;

[0180] Displacement fault feature: Monitor the displacement signal fluctuations through the time-domain and frequency-domain characteristics of the displacement signal;

[0181] Signal change rate: Monitor the sharp changes in the signal.

[0182] Construct a random forest model based on historical experience, and input the fault feature set into the random forest model; Each decision tree in the random forest is trained based on different subsets of the fault feature set to form independent classification rules for identifying different fault modes;

[0183] Each decision tree in the random forest independently classifies and predicts the fault feature set, and each tree predicts the fault type corresponding to the current feature according to its classification rules;

[0184] The random forest uses a voting mechanism to make a final judgment on the fault type and finally generates a fault identification result.

[0185] Implement an adaptive model update mechanism to adjust and update the fault identification model;

[0186] The steps of the adaptive update mechanism are as follows: regularly collect the latest operation data of the pneumatic control valve, and generate a real-time fault feature set according to the known feature extraction rules;

[0187] Use an adaptive online learning algorithm to update the parameters of the fault identification model according to the real-time fault feature set; make real-time adjustments to the parameters of the fault identification model and update the model through online learning;

[0188] The loss function reflects the model classification error. Calculate the loss function according to the difference between the real-time data and the identification result; when the loss function increases, the adaptive mechanism will be automatically triggered, and the fault identification model will be adjusted by introducing new feature data.

[0189] Collect the historical operation data of the pneumatic control valve, and use a deep learning model to train and optimize the fault identification model to obtain a multi-fault identification model.

[0190] The training process of the multi-fault identification model is as follows:

[0191] Collect the historical operation data of the pneumatic control valve, including the historical feature data and the corresponding fault labels of the feature data output by the fault identification model; after preprocessing, input it into the deep neural network to extract the relationship features between the operation data and the fault type;

[0192] The optimization process of the multi-fault identification model is as follows:

[0193] Use the cross-entropy loss function to optimize the preliminary multi-fault identification model;

[0194] Process the data collected from Chemical Plant A in Table 2 according to the above steps to obtain the multi-fault types of the pneumatic control valve operation system as air chamber leakage, valve stem jamming or actuator abnormality, valve overheating, and mechanical looseness.

[0195] In this embodiment, through the preprocessing steps, outlier detection, low-pass filtering, and baseline correction are performed on the collected air chamber pressure, valve stem displacement, temperature, and vibration signals to ensure the quality and accuracy of the data, laying a foundation for subsequent analysis; then, using edge computing and lightweight data processing technologies, the data can be quickly analyzed and processed, reducing the computing load and space occupancy, providing data support for improving the real-time performance and efficiency of fault detection; then, through the feature extraction of multi-source data signals, the operation state of the pneumatic control valve equipment can be comprehensively captured, and multiple fault types of the pneumatic control valve can be identified.

[0196] In the fault identification stage, in this embodiment, through deep learning and random forest models, complex concurrent faults are accurately classified and identified to ensure efficient processing of different types of fault signals; through an adaptive mechanism, the fault identification model can automatically adjust model parameters according to changes in actual operation data to ensure robustness during long-term operation and continuous optimization of fault identification.

[0197] This embodiment provides data, theory, and technical support for the real-time online monitoring of pneumatic control valves through multiple steps such as lightweight processing, edge computing, key feature extraction, and fault identification, improving the accuracy, real-time performance, and robustness of the real-time online monitoring of pneumatic control valves.

[0198] Although the embodiments of the present invention have been shown and described, for those of ordinary skill in the art, it can be understood that various changes, modifications, substitutions, and variations can be made to these embodiments without departing from the principles and spirit of the present invention, and the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A real-time online monitoring method for a pneumatic control valve, characterized in that: include: Collect real-time signal data during the operation of the pneumatic control valve and transmit it to the edge computing node for processing; The real-time signal data includes air chamber pressure data, valve stem displacement data, temperature data and vibration data; Preprocessing and lightweight processing are performed on the real-time signal data to generate frequency spectrum data of the regulating valve; Extracting key frequency components from the frequency spectrum data of the regulating valve by feature calculation, generating regulating valve feature data, and transmitting the data to a central computing node; The process of acquiring the characteristic data of the regulating valve includes: Extract the maximum frequency corresponding to the highest amplitude from the frequency spectrum data of the regulating valve; calculate the power spectrum density of the frequency spectrum data of the regulating valve; calculate the root mean square value of the frequency spectrum data of the regulating valve; perform abnormal monitoring on the maximum frequency and the root mean square value, and mark them as abnormal when the detected maximum frequency and the root mean square value are beyond the normal range; compress the maximum frequency, the root mean square value and the power spectrum density using a compression function to obtain the characteristic data of the regulating valve, and transmit them to the central computing node; Extracting and reducing the dimension of the regulating valve characteristic data to obtain the principal component characteristics; Performing multi-source data fusion on the principal component features to generate principal component fusion features; F=ω P ·Z P +ω d ·Z d +ω T ·Z T +ω a ·Z a ; Among them, Z P , Z d , Z T and Z a are the principal component characteristics of pressure, displacement, temperature and vibration, ω P ,ω d ,ω T and ω a are the weights of the corresponding principal component features; Analyzing the fusion features and extracting key features to generate a control valve fault feature set; The random forest model is used to classify and identify the control valve fault feature set to obtain the fault identification result; Adopting an adaptive model update mechanism to adjust and update the random forest model; Collecting historical operation data of the pneumatic control valve and the fault identification results, performing extended training and optimization on the random forest model, and obtaining a multi-fault identification model; The training process of the multi-fault identification model is as follows: collecting historical data of the pneumatic control valve, including historical feature data and the corresponding fault identification results; performing the extended training on the random forest model to extract the correlation features between the historical feature data and the fault identification results; adjusting the random forest model through the correlation features to obtain the multi-fault identification model; The optimization process of the multi-fault identification model is: The latest operating data of the pneumatic control valve is collected, and the multi-fault identification model is used to obtain the multi-fault identification result; the cross entropy loss function is used to calculate the difference between the multi-fault identification result and the actual multi-fault type in real time, and the multi-fault identification model is optimized and adjusted according to the difference.

2. A real-time online monitoring method for a pneumatic control valve according to claim 1, characterized in that: The preprocessing includes low-pass filtering, baseline correction and outlier detection filtering; Removing high-frequency noise from the real-time signal data by using a low-pass filter; The moving average method is used to eliminate the baseline drift in the data signal and correct the data signal; Identify and exclude outliers in data signals using the three-times standard deviation method.

3. The real-time online monitoring method of a pneumatic control valve according to claim 1 is characterized in that: The generation process of the regulating valve spectrum data is as follows: The pre-processed real-time signal data is converted into a frequency domain signal by fast Fourier transform to generate the regulating valve spectrum data.

4. The real-time online monitoring method of a pneumatic control valve according to claim 1 is characterized in that: The multi-source data fusion step includes principal component analysis and weighted fusion; The regulating valve characteristic data includes air chamber pressure characteristics, valve stem displacement characteristics, temperature characteristics and vibration characteristics; Extracting principal components from the regulating valve characteristic data using principal component analysis and reducing the dimension to generate principal component features; The principal component features are weighted and fused to generate principal component fusion features.

5. The real-time online monitoring method of a pneumatic control valve according to claim 1, characterized in that: The steps of obtaining the fault feature set are: Extract key features related to the fault from the principal component fusion features through a feature extraction function to generate a fault feature set; The fault feature set includes vibration fault features, pressure fault features, temperature fault features, displacement fault features and signal change rate.

6. A real-time online monitoring method for a pneumatic control valve according to claim 1, characterized in that: The step of obtaining the fault identification result includes: Constructing a random forest model, and inputting the fault feature set into the random forest model; The fault feature set is divided into several training subsets for training, and independent classification rules are formed for the fault modes; Classify and predict the fault feature set; wherein each tree in the random forest predicts the fault type corresponding to the current feature according to the corresponding independent classification rule; The fault identification result is obtained by adopting a majority vote mechanism.

7. A real-time online monitoring method for a pneumatic control valve according to claim 1, characterized in that: The steps of the adaptive update mechanism are: Calculating the difference between the fault identification result and the actual fault type to determine the triggering condition of the adaptive mechanism; Collect the latest operating data of the pneumatic control valve to generate a real-time fault feature set; The parameters of the random forest model are updated using an adaptive online learning algorithm according to the real-time fault feature set.

8. The real-time online monitoring method of a pneumatic control valve according to claim 1, characterized in that: The extended training process includes: Input the historical feature data and the corresponding fault identification result into a three-layer convolution layer, extract fault correlation features, and output a first correlation feature map; Using maximum pooling to reduce the dimension of the first correlation feature map, reducing the size of the first correlation feature map and retaining important information, and outputting a second correlation feature map; Expand the second correlation feature map into a vector and extract the fault correlation global feature; The fault-associated global features are converted into probability distribution of fault types through a softmax activation function to obtain the associated features.

Images