A pneumatic control valve fault diagnosis system based on cloud theory

Abstract

The present invention discloses a pneumatic control valve fault diagnosis system based on cloud theory, including a fault signal acquisition module, a cloud feature extraction module, a dynamic feature extraction module, a fault feature fusion module, a cloud sampling module, a base model training module, and an integrated decision diagnosis module. Based on the above modules, the present invention counts the diagnosis results of each incremental random weight neural network base model under the same test sample, and takes the mode of the diagnosis results as the final diagnosis result of the integrated model. The present invention not only realizes the fault diagnosis of the pneumatic control valve with a high diagnostic accuracy, and effectively avoids the operation of the valve with a fault; and considering the problem of the lack of valve fault data in practical applications, the problem of low diagnostic accuracy of the model under small sample data is solved by constructing samples and integration methods. In addition, the present invention can complete fault diagnosis without being familiar with the valve mechanism and complicated empirical knowledge, and has good practicality.

Description

Technical Field

[0001] The present invention relates to the field of valve fault detection and diagnosis, and in particular to a pneumatic control valve fault diagnosis system based on cloud theory. Background Art

[0002] As the basic control unit in the industrial fluid delivery system, the industrial control valve is a vital part to ensure the safety of industrial production and the smooth operation of the system. It has the advantages of stable performance and rapid action. However, due to frequent mechanical movements, harsh industrial production environment and strong corrosiveness of the transport medium, the control valve is prone to various abnormal faults. These faults may affect industrial production at the least and threaten people's lives and property at the worst. Therefore, improving the fault diagnosis ability of the control valve is of great significance to ensure the safe and stable operation of the industrial fluid delivery system.

[0003] At present, the research on fault diagnosis methods is mainly divided into three aspects: methods based on analytical models, methods based on expert knowledge and methods based on data-driven.

[0004] The diagnostic method based on analytical model is to analyze the input and output characteristics of the control valve and establish a mathematical model for fault diagnosis based on a full understanding of the mechanism and structure of the control valve, combined with relevant knowledge such as mass flow and dynamics theory, aeroacoustics and vortex motion theory, pressure field and turbulent jet theory. The advantage of this method is that it can intuitively reflect the characteristics of the fault from the mathematical model, but due to the differences in input quantities and interference factors under different working conditions, it is often difficult to establish an accurate and universal analytical model, so there are great limitations in practical applications.

[0005] Expert knowledge-based diagnostic methods use expert knowledge bases to infer and diagnose faults. Such methods include expert systems, fuzzy logic, etc. Since expert rules are successful experiences summarized and mined from the system operation process, they have higher practical application value, but their problems are also obvious: the formulation of rules depends on human experience, the learning ability of the model is limited, the ability to handle unknown situations is poor, and when there are a large number of expert rules, the rule search will inevitably reduce the response speed of the system.

[0006] In recent years, with the rapid rise of various intelligent algorithms, data-driven methods have attracted widespread attention from scholars. Such methods mainly include statistical analysis and artificial intelligence methods. Statistical analysis methods include control chart method, principal component analysis method, partial least squares method, etc. Artificial intelligence methods include neural network, support vector machine, extreme learning machine, etc. Its main idea is to use process data training to obtain a model of specific parameters of the diagnosed object, thereby achieving the purpose of process monitoring. Its advantage is that it is based on process data and does not require the establishment of complex analytical models. At the same time, such methods do not rely on expert knowledge, and the established model has stronger learning ability. However, many data-driven valve fault diagnosis methods also have some problems.

[0007] First, there is a lack of effective extraction and utilization of the uncertainty components in the fault signal. Due to the interference of the external environment and noise, the influence of data acquisition and information processing, etc., the fault information of the control valve is uncertain. Many current studies only perform simple mathematical transformations on the data or generate residual signals indicating faults without considering the uncertainty of the fault information, which may affect the accuracy of the diagnosis model.

[0008] Second, the learning algorithm modeling process used to establish the diagnostic model is too time-consuming and difficult to reconstruct online. Taking the BP neural network algorithm as an example, in the flat area of ​​the error gradient surface, the weight error changes very little, making the algorithm converge slowly.

[0009] Third, there is a lack of fault data in actual industrial applications. This problem is particularly prominent in the context of control valves, and it is difficult to collect enough data to meet the needs of model training. Without sufficient training samples to train the model, the diagnostic accuracy of the learning algorithm may be low, that is, the training result is a weak learning model. Summary of the invention

[0010] The technical problem to be solved by the present invention is to overcome the defects of the prior art and provide a pneumatic control valve fault diagnosis system based on cloud theory with simple structure and obvious effect.

[0011] The purpose of the present invention is mainly achieved through the following technical solutions:

[0012] The present invention provides a pneumatic control valve fault diagnosis system based on cloud theory, including a fault signal acquisition module, a cloud feature extraction module, a dynamic feature extraction module, a fault feature fusion module, a cloud sampling module, a base model training module, and an integrated decision diagnosis module;

[0013] The fault signal acquisition module is used to obtain the time series of the control valve fault signal by means of a step-by-step test and pre-process it;

[0014] The cloud feature extraction module is used to extract fault cloud features from the time series of fault signals according to cloud model theory;

[0015] The dynamic feature extraction module is used to extract l dynamic features from the time series of the fault signal according to the dynamic internal principal component analysis method;

[0016] The fault feature fusion module is used to fuse the fault cloud features and dynamic features to obtain a small sample training set, denoted as X;

[0017] The cloud sampling module is used to mine cloud feature information from a small sample training set and map it to the cloud feature space to construct training samples, thereby achieving a data enhancement effect;

[0018] The base model training module is used to train different control valve fault diagnosis base models based on the generated fault cloud droplets using an incremental random weight neural network algorithm;

[0019] The integrated decision-making diagnosis module is used to count the diagnosis results of each incremental random weight neural network base model under the same test sample, and take the mode of the diagnosis results as the final diagnosis result of the integrated model.

[0020] Furthermore, the execution process of the fault signal acquisition module described in the present invention includes: there are N states of the pneumatic control valve, including 1 normal state and N-1 fault states; in the process of collecting the time series of the fault signal, each time the controller gives an input signal CV starting from 0% (the controller belongs to the hardware of the control valve itself, and the valve opening is determined by the signal size given by the controller), and the σ% of the total allowable movement stroke of the valve stem is used as the difference change for graded testing until CV reaches 100%, and a total of N1 different opening states are experienced; the time series of the fault signal collected containing N1 different opening states is normalized.

[0021] The cloud feature extraction module specifically performs the following steps: taking the valve stem displacement and medium flow rate data obtained by the fault signal acquisition module under a fault state, and extracting the cloud features of the control valve under different opening states based on the inverse cloud algorithm of cloud theory, namely Ex, En, and He, where Ex is the value of the fault signal that best represents the concept of the control valve fault under the current opening state, En represents the uncertainty of the control valve fault signal, and He is a hyperparameter that represents the uncertainty of En. The two fault signals of valve stem displacement and medium flow rate can be used to extract N1×3×2-dimensional control valve fault cloud feature values ​​through the same steps.

[0022] The cloud feature extraction module also performs the following steps:

[0023] Step a1: Assume that the sampling sequence of a fault signal obtained by the fault signal acquisition module when the pneumatic control valve is in a fault state M is a=[a1, a2, ..., a i ], where a i is composed of n elements a i (k), k=1,2,…,n composed of a column vector, a i (k) represents the kth fault information cloud droplet of the regulating valve at the i-th opening;

[0024] Step a2: find the sampling sequence a of the fault signal at the i-th opening i The mean expect

[0025] Step a3, calculate column vector a i The first central moment of entropy

[0026] Step a4, calculate column vector a i Variance Super Entropy

[0027] Step a5, obtain the cloud model feature sequence B = [B1, B2, ..., B i ], where B i =[Ex i ,En i ,He i ], B i Represents the cloud model eigenvalue of the control valve at the i-th opening.

[0028] Let Q = [q1 q2 … q z+s ] T is z+s fault signal data under a fault state (parameter z has no specific meaning, written as z+s to facilitate the expression of subsequent steps), where s represents the time order, q z+s Represents the time series of the z+sth fault signal.

[0029] The dynamic feature extraction module specifically performs the following steps:

[0030] Step b1, initializing the weight vector ζ to a random unit vector;

[0031] Step b2, extract the main element c = Qζ, take the tth to t+z-1th rows from c, denoted as ψ t , where t=1,2,…,s+1;

[0032] Step b3, update the autoregressive model parameters η = [ψ1 ψ2 … ψ s ]T ψ s+1 and weight vector And unitized respectively, where η t represents the tth element of the column vector η of length s, Q t =[q t q t+1 … q z+t-1 ], T represents the transpose operation;

[0033] Step b4, save the current main element c, update the sample data Q = Q-cp T , where p = Q T c / c T c represents the load vector after extracting the principal element c;

[0034] Step b5, extract the next pivot element, until l pivot elements are extracted;

[0035] Step b6, obtain the dynamic feature sequence C = [c1, c2, ..., c l ], c l Represents the extracted lth pivot.

[0036] The fault feature fusion module reads B and C respectively, and X=[BC], then X is the fault feature sequence of the regulating valve after the fault state M is fused by the fault feature fusion module, with a total of N1×3×2+1 dimensions. Since the number of feature sequences actually obtained is small, the present invention also calls X a small sample training set.

[0037] The cloud sampling module specifically performs the following steps:

[0038] Step c1, setting the number of cloud droplets generated to represent the control valve fault information, that is, the number of training samples d required for each fault state of each control valve fault diagnosis base model and the threshold value θ of the corresponding certainty of the generated cloud droplets;

[0039] Step c2: Take all small sample training sets X corresponding to the fault state M and calculate the cloud model feature value Ex of the j-th dimension feature Feature j under the fault state M. j 、En j 、He j , where Ex j represents the expectation of the j-th dimension feature, En j The cloud feature entropy He represents the j-th dimension feature. j represents the super entropy of the j-th dimension feature, j = 1, 2, ..., N1 × 3 × 2 + l;

[0040] Step c3, generating a regulating valve fault cloud droplet according to the cloud model characteristic value;

[0041] Step c4, calculating the certainty of the qualitative concept belonging to the fault state M corresponding to the generated regulating valve fault cloud droplet;

[0042] Step c5, when the degree of certainty corresponding to the regulating valve fault cloud droplet is greater than the set threshold value θ, it is retained, otherwise it is regenerated until d regulating valve fault cloud droplets that meet the conditions are generated;

[0043] Step c6, merging the control valve fault cloud droplets generated by each dimension of features and labeling them to form a training sample of fault M;

[0044] Step c7, repeat steps c1 to c6 until N training samples of fault states are generated and combined, thus completing the construction of an incremental random weight neural network base model training sample. This process is called cloud sampling.

[0045] The base model training module is trained on the fault cloud droplets generated by the cloud sampling module to obtain a total of N2 incremental random weight neural network base models, and the output Y is obtained. out =H*β L , where H is the hidden layer output matrix, β L To output the weight, the Sigmoid function is selected as the activation function. The integrated decision diagnosis module counts the test results of each incremental random weight neural network base model for each test sample, and selects the mode as the final diagnosis result of the corresponding test sample.

[0046] Beneficial effects: The present invention has a simple structure and obvious effects. It not only realizes the fault diagnosis of pneumatic control valves with a high accuracy, but also reduces the rate of missed fault diagnosis, effectively avoiding the operation of valves with faults. Moreover, considering the lack of valve fault data in practical applications, the problem of low diagnostic accuracy of the model under small sample data is solved by constructing samples and integrating methods. In addition, fault diagnosis can be completed without being familiar with valve mechanisms and complicated empirical knowledge, which has good practicality and is of certain significance for promoting the intelligent development of control valve fault diagnosis. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] The present invention will be further described in detail below in conjunction with the accompanying drawings and specific embodiments, and the above and other advantages of the present invention will become more clear.

[0048] Figure 1 It is a schematic diagram of the system structure of the present invention;

[0049] Figure 2 A pneumatic regulating valve used in the present invention;

[0050] Figure 3 This is a schematic diagram of cloud sampling;

[0051] Figure 4The test confusion matrix of the diagnosis result of the case shown in the present invention, the sum of each row represents the number of real samples of this category, and the sum of each column represents the number of samples predicted to be of this category. DETAILED DESCRIPTION

[0052] The present invention will be further described below in conjunction with the accompanying drawings and embodiments.

[0053] Example

[0054] like Figure 1 As shown, the present invention discloses a pneumatic control valve fault diagnosis system based on cloud theory, including a fault signal acquisition module, a cloud feature extraction module, a dynamic feature extraction module, a fault feature fusion module, a cloud sampling module, a base model training module, and an integrated decision diagnosis module: the fault signal acquisition module is used to obtain the time series of the control valve fault signal by a step test method and pre-process it; the cloud feature extraction module is used to extract the fault cloud feature from the time series of the fault signal obtained by the fault signal acquisition module according to the cloud model theory; the dynamic feature extraction module is used to extract the fault dynamic feature from the time series of the fault signal obtained by the fault signal acquisition module by a dynamic internal principal component analysis method; the fault feature fusion module is used to fuse the extracted fault cloud feature and dynamic feature; the cloud sampling module is used to construct training samples from the cloud feature space; the base model training module is used to train different control valve fault diagnosis base models based on the generated fault cloud droplets by using an incremental random weight neural network algorithm; the integrated decision diagnosis module is used to count the diagnosis results of each incremental random weight neural network base model under the same test sample, and take the mode as the final diagnosis result of the integrated model.

[0055] A specific embodiment of the present invention, combined with Figure 2 , is the pneumatic control valve used in the present invention; in the figure: 1 spring 2 air chamber 3 controller 4 valve stem 5 valve seat. First, the fault signal acquisition module of the present invention adopts a graded test method to obtain the fault signal, and its specific process includes: there are 10 states of the pneumatic control valve, including 1 normal state and 9 fault states; in the process of training the diagnosis model, each time the controller gives an input signal CV starting from 0%, 25% of the total stroke allowed by the valve stem is used as the difference change for graded testing until CV reaches 100%, and a total of 5 different opening states are experienced, and the valve stem displacement and medium flow rate collected in 5 different opening states are normalized.

[0056] Secondly, the cloud feature extraction module and the dynamic feature extraction module respectively extract features from the acquired fault signal, and the cloud feature extraction module further performs the following steps:

[0057] Step a1: Assume that the sampling sequence of a certain fault signal obtained by the fault signal acquisition module under a certain fault state M of the pneumatic control valve is a=[a1, a2, ..., a i ], where a i is composed of n elements a i (k), k=1,2,…,n composed of a column vector, each a i (k) is called a fault information cloud drop, i represents different controller output states, i.e., different openings of the regulating valve;

[0058] Step a2: find the sampling sequence a of the fault signal at the i-th opening i The mean expect

[0059] Step a3, calculate the fault signal sampling sequence a i The first central moment of entropy

[0060] Step a4, calculate the fault signal sampling sequence a i variance Super Entropy

[0061] Step a5, let B = [B1, B2, ..., B i ], where B i =[Ex i ,En i ,He i ], represents the cloud model eigenvalue of the control valve at the i-th opening. The cloud model feature sequence B under the fault state M is obtained through the above steps. Ex is the value of the fault signal that best represents the concept of the control valve fault under the current opening state. En represents the uncertainty of the control valve fault signal. He is a hyperparameter that represents the uncertainty of En. Then, 5 groups of cloud model eigenvalues ​​are extracted from the fault signal. The same steps are used to extract 30 dimensions of cloud model eigenvalues ​​from the two fault signals.

[0062] Let Q = [q1 q2 … q z+s ] T is z+s fault signal data under a fault state, where s represents the time order;

[0063] The specific execution steps of the dynamic feature extraction module include:

[0064] Step b1, initializing the weight vector ζ to a random unit vector;

[0065] Step b2, extract the main element c = Qζ, take the tth to t+z-1th rows from c, denoted as ψ t, where t=1,2,…,s+1;

[0066] Step b3, update the autoregressive model parameters η = [ψ1 ψ2 … ψ s ] T ψ s+1 and weight vector And unitized respectively, where η t represents the tth element of the column vector η of length s, Q t =[q t q t+1 … q z+t-1 ], T represents the transpose operation;

[0067] Step b4, save the current main element c, update the sample data Q = Q-cp T , where p = Q T c / c T c represents the load vector after extracting the principal element c;

[0068] Step b5, extract the next pivot element, until l pivot elements are extracted;

[0069] Step b6, obtain the dynamic feature sequence C = [c1, c2, ..., c l ], c l represents the extracted lth pivot;

[0070] Furthermore, the fault feature fusion module reads B and C respectively, and X=[BC], then X is the fault feature sequence of the regulating valve after the fault state M is fused by the fault feature fusion module, with a total of 30+1 dimensions. Since the number of feature sequences obtained is small, the present invention also calls X a small sample training set.

[0071] A specific embodiment of the present invention, as Figure 3 FIG. 1 is a schematic diagram of the cloud sampling process of the present invention. The execution steps of the cloud sampling module include:

[0072] Step c1, setting the number of cloud droplets generated to represent the control valve fault information, that is, the number of training samples d required for each fault state of each control valve fault diagnosis base model is set to 140 and the threshold value θ of the degree of certainty corresponding to the generated cloud droplets is set to 0.9;

[0073] Step c2: Take all small sample training sets X corresponding to the fault state M, and calculate the cloud model feature value Ex of the j-th dimension feature Feature j under the fault state M according to the method described in the cloud feature extraction module. j 、En j 、He j , where Ex j represents the expectation of the j-th dimension feature, En jThe cloud feature entropy He represents the j-th dimension feature. j represents the super entropy of the j-th dimension feature, j = 1, 2, ..., 30 + l;

[0074] Step c3, characterize En with the uncertainty of the control valve fault signal j is the expectation, with the hyperparameter He j Generate normal random numbers En' for the standard deviation j ; The value Ex of the fault signal that characterizes the fault concept of the control valve j is the expectation, with the generated random number En' j Generate control valve fault characteristic cloud droplet x for standard deviation j ;

[0075] Step c4, calculate x j The corresponding degree of certainty of the qualitative concept belonging to the fault state M

[0076] Step c5, if the certainty corresponding to the generated regulating valve fault characteristic cloud droplet is greater than the set threshold value θ (value 0.9), it is retained, otherwise it is regenerated until d (value 140) cloud droplets that meet the conditions are generated;

[0077] Step c6, merging the cloud droplets generated by each dimension of features and labeling them to form a training sample of fault M;

[0078] Step c7, loop steps c1 to c6 until N (value is 10) training samples of fault states are generated and combined, thus completing the construction of an incremental random weight neural network base model training sample. This process is called cloud sampling.

[0079] Furthermore, the base model training module trains a total of N2 (value is 3) incremental random weight neural network base models based on the fault cloud droplets generated by the cloud sampling module. The modeling process of each base model needs to re-execute the cloud sampling process to obtain different training samples, so that each base model has discrimination.

[0080] It should be noted that the modeling process of the incremental random weighted neural network base model of the present invention includes the following steps:

[0081] Step d1, initialization parameters. Maximum number of hidden layer nodes L max =50, critical tolerance error ε = 0.15, upper limit of random parameter change λ = 1;

[0082] Step d2, randomly configure the hidden layer parameters. Use random numbers to configure the input weight ω and bias b, where ω = r1×2λ-λ, b = r2×2λ-λ, r1 is a random sequence that follows a uniform distribution, and r2 is a random number that follows a uniform distribution;

[0083] Step d3, calculate the output matrix h L , the activation function selects the Sigmoid function, and the calculation formula is:

[0084] Step d4, calculate the hidden layer output weight Among them, H L =[h1,h2,…,h L ],h L represents the output matrix of the Lth node, represents the generalized inverse, and Y is the output matrix generated by one-hot encoding of the sample labels. The residual is e L =H L β L -Y, when the residual meets the tolerance error ε or the number of hidden layer nodes reaches L max No more nodes are added when the modeling is completed, and an incremental random weight neural network base model f = H is obtained. L β L ;

[0085] Step d5, repeat steps d1 to d4 until three incremental random weight neural network base models are obtained.

[0086] Input the same test sample into the base model training module to obtain the three incremental random weight neural network base models, and obtain their output Y out =H*β L , where H is the hidden layer output matrix, β L The Sigmoid function is selected as the output weight, and the integrated decision diagnosis module counts the test results of each base model for each test sample, and selects the mode as the final diagnosis result of the corresponding test sample; then the diagnostic accuracy is calculated by comparing with the test sample label. The calculation formula is: Where N F Represents the number of label classification errors in the test sample, N A Represents the total number of all test samples. The confusion matrix can be used to visually observe the model diagnosis results. The diagnosis result confusion matrix of this example is as follows Figure 4 As shown, the diagnostic accuracy rate reached 97.67%.

[0087] The present invention provides a pneumatic control valve fault diagnosis system based on cloud theory. There are many methods and ways to implement the technical solution. The above is only a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principle of the present invention. These improvements and modifications should also be regarded as the protection scope of the present invention. All components not specified in this embodiment can be implemented by existing technologies.

Claims

1. A pneumatic control valve fault diagnosis system based on cloud theory, characterized in that: It includes fault signal acquisition module, cloud feature extraction module, dynamic feature extraction module, fault feature fusion module, cloud sampling module, base model training module, and integrated decision diagnosis module; The fault signal acquisition module is used to obtain the time series of the control valve fault signal by means of a step-by-step test and pre-process it; The cloud feature extraction module is used to extract fault cloud features from the time series of fault signals according to cloud model theory; The dynamic feature extraction module is used to extract l dynamic features from the time series of the fault signal according to the dynamic internal principal component analysis method; The fault feature fusion module is used to fuse the fault cloud features and dynamic features to obtain a small sample training set, denoted as X; The cloud sampling module is used to mine cloud feature information from a small sample training set and map it to the cloud feature space to construct training samples, thereby achieving a data enhancement effect; The base model training module is used to train different control valve fault diagnosis base models based on the generated fault cloud droplets using an incremental random weight neural network algorithm; The integrated decision-making diagnosis module is used to count the diagnosis results of each incremental random weight neural network base model under the same test sample, and take the mode of the diagnosis results as the final diagnosis result of the integrated model; The cloud sampling module specifically performs the following steps: Step c1, setting the number of cloud droplets generated to represent the control valve fault information, that is, the number of training samples d required for each fault state of each control valve fault diagnosis base model and the threshold value θ of the corresponding certainty of the generated cloud droplets; Step c2: Take all small sample training sets X corresponding to the fault state M and calculate the cloud model feature value Ex of the j-th dimension feature Featurej under the fault state M. j 、En j 、He j , where Ex j represents the expectation of the j-th dimension feature, En j The cloud feature entropy He represents the j-th dimension feature. j represents the super entropy of the j-th dimension feature, j = 1, 2, ..., N1 × 3 × 2 + l; Step c3, generating a regulating valve fault cloud droplet according to the cloud model characteristic value; Step c4, calculating the certainty of the qualitative concept belonging to the fault state M corresponding to the generated regulating valve fault cloud droplet; Step c5, when the degree of certainty corresponding to the regulating valve fault cloud droplet is greater than the set threshold value θ, it is retained, otherwise it is regenerated until d regulating valve fault cloud droplets that meet the conditions are generated; Step c6, merging the control valve fault cloud droplets generated by each dimension of features and labeling them to form a training sample of fault M; Step c7, repeat steps c1 to c6 until N training samples of fault states are generated and combined, thus completing the construction of an incremental random weight neural network base model training sample. This process is called cloud sampling.

2. The system according to claim 1, characterized in that The fault signal acquisition module specifically performs the following steps: there are N pneumatic control valve states, including 1 normal state and N-1 fault states; in the process of acquiring the time series of the fault signal, each time the controller gives an input signal CV starting from 0%, a graded test is performed with σ% of the total allowable movement stroke of the valve stem as the difference change until CV reaches 100%, and a total of N1 different opening states are experienced; the time series of the acquired fault signal containing N1 different opening states is normalized.

3. The system according to claim 2, characterized in that The cloud feature extraction module specifically performs the following steps: taking the valve stem displacement and medium flow rate data obtained by the fault signal acquisition module under a fault state, and extracting the cloud features of the control valve under different opening states based on the inverse cloud algorithm of cloud theory, namely Ex, En, and He, where Ex is the value of the fault signal that best represents the concept of the control valve fault under the current opening state, En represents the uncertainty of the control valve fault signal, and He is a hyperparameter that represents the uncertainty of En. The two fault signals of valve stem displacement and medium flow rate can be used to extract N1×3×2-dimensional control valve fault cloud feature values ​​through the same steps.

4. The system according to claim 3, characterized in that The cloud feature extraction module also performs the following steps: Step a1: Assume that the sampling sequence of a fault signal obtained by the fault signal acquisition module when the pneumatic control valve is in a fault state M is a=[a1, a2, ..., a i ], where a i is composed of n elements a i (k), k=1,2,…,n composed of a column vector, a i (k) represents the kth fault information cloud droplet of the regulating valve at the i-th opening; Step a2: find the sampling sequence a of the fault signal at the i-th opening i The mean expect Step a3, calculate column vector a i The first central moment of entropy Step a4, calculate column vector a i Variance Super Entropy Step a5, obtain the cloud model feature sequence B = [B1, B2, ..., B i ], where B i =[Ex i ,En i ,He i ], B i Represents the cloud model eigenvalue of the control valve at the i-th opening.

5. The system according to claim 4, characterized in that Q=[q1 q2…q z+s ] T is z+s fault signal data under a fault state, where s represents the time order, q z+s Represents the time series of the z+sth fault signal.

6. The system according to claim 5, characterized in that The dynamic feature extraction module specifically performs the following steps: Step b1, initializing the weight vector ζ to a random unit vector; Step b2, extract the main element c = Qζ, take the tth to t+z-1th rows from c, denoted as ψ t , where t=1,2,…,s+1; Step b3, update the autoregressive model parameters η = [ψ1ψ2…ψ s ] T ψ s+1 and weight vector And unitized respectively, where η t represents the tth element of the column vector η of length s, Q t =[q t q t+1 …q z+t-1 ], T represents the transpose operation; Step b4, save the current main element c, update the sample data Q = Q-cp T , where p = Q T c / c T c represents the load vector after extracting the principal element c; Step b5, extract the next pivot element, until l pivot elements are extracted; Step b6, obtain the dynamic feature sequence C = [c1, c2, ..., c l ], c l Represents the extracted lth pivot.

7. The system according to claim 6, characterized in that The fault feature fusion module reads B and C respectively, and X=[BC], then X is the fault feature sequence of the regulating valve after the fault state M is fused by the fault feature fusion module, with a total of N1×3×2+1 dimensions, and X is also called a small sample training set.

8. The system according to claim 7, characterized in that The base model training module is trained on the fault cloud droplets generated by the cloud sampling module to obtain a total of N2 incremental random weight neural network base models, and the output Y is obtained. out =H*β L , where H is the hidden layer output matrix, β L To output the weight, the Sigmoid function is selected as the activation function. The integrated decision diagnosis module counts the test results of each incremental random weight neural network base model for each test sample, and selects the mode as the final diagnosis result of the corresponding test sample.

Images