CN112229624A - Pneumatic regulating valve fault diagnosis method based on low-deviation random configuration network - Google Patents

Abstract

The invention discloses a pneumatic regulating valve fault diagnosis method based on a low-deviation random configuration network. The invention not only realizes the fault diagnosis of the pneumatic regulating valve, but also has higher diagnosis accuracy, and effectively avoids the fault operation of the valve; the method has good universality, and fault diagnosis can be completed without being familiar with valve mechanism and complicated experience knowledge; general operating personnel can master, has improved pneumatic governing valve fault diagnosis's automation and intelligent degree.

Description

Pneumatic regulating valve fault diagnosis method based on low-deviation random configuration network

Technical Field

The invention relates to the field of valve fault detection and diagnosis, in particular to a pneumatic regulating valve fault diagnosis method based on a low-deviation random configuration network.

Background

The regulating valve is an important component of a fluid control system, and plays important roles in controlling a flow direction, managing opening and closing, regulating medium parameters and the like in the system. In order to operate stably, the valve should be maintained and maintained regularly to prolong the service life and reliability, but the valve inevitably fails in a long-term production process, thereby affecting the production practice. Under the huge industrial scale of China, the research on the fault diagnosis of the regulating valve has important significance.

The current research of fault diagnosis methods is mainly divided into three aspects: analytical model-based methods, qualitative model-based methods, and data-driven-based methods.

The method based on the analytical model needs to accurately test the dynamic characteristics of the process object, and obtains a mathematical model of the object by applying methods such as priori knowledge, system identification and the like, so that the mathematical model can be described by an arithmetic equation and an integral differential equation. When the system is complex, it is difficult to build a quantitative model of the system. For the valve fault diagnosis system, even if the valve is a pneumatic valve, different manufacturers make respective innovation under a general framework, so that the model is difficult to have universality.

The method based on the qualitative model does not need to accurately describe the object, only needs to establish the qualitative model of the object, and describes the structure of the system by using the fixed variables representing the physical parameters and the qualitative constraint equation representing the mutual relation between the parameters, but only uses the qualitative information, so that the problems of redundancy, inconsistent diagnosis conclusion and the like of the qualitative reasoning result are easily caused. Meanwhile, the monitoring method based on qualitative experience requires many complicated and profound professional knowledge and long-term experience accumulation, so that the operation is difficult.

Data-driven fault diagnosis methods have attracted attention of many scholars in recent years, mainly by statistical analysis methods and artificial intelligence methods. The statistical analysis method includes a control graph method, a principal component analysis method, a partial least square method and the like. The artificial intelligence method comprises a neural network, a support vector machine, an ELM and the like, and the basic idea is to use process data training to obtain a model of specific parameters of a diagnosed object so as to achieve the purpose of process monitoring. Its advantages are no need of defining the physical law and structure principle of object to be measured. Data-driven fault diagnosis based methods have attracted attention of many scholars in recent years, and randomization algorithms for large-scale data modeling have become popular for research. Some popular random networks have some problems.

RWNNs (Random weight neural networks) are simple to implement and fast in modeling speed. However, two essential drawbacks have limited the use of RWNNs: 1) the number of hidden layer nodes cannot be determined prior to training, i.e. it is difficult to set a suitable network structure; 2) hidden layer parameters are generated in a fixed and unchangeable interval, and the actual approximation characteristics of the hidden layer parameters are influenced.

The RVFLNs (Random vector function link networks) randomly distribute input weight and bias, and the least square method is adopted to train output weight, so that the model training speed is higher. However, if the range of the hidden layer random parameter settings is not appropriate, the RVFLNs network cannot approach the objective function with probability 1.

Disclosure of Invention

The invention aims to overcome the defects of the prior art and provides a pneumatic regulating valve fault diagnosis method based on a low-deviation random configuration network, which has a simple structure and an obvious effect.

In order to solve the technical problem, the pneumatic regulating valve fault diagnosis method based on the low-deviation random configuration network provided by the invention comprises the following steps of:

1. data acquisition:

a displacement sensor is arranged at a reserved position of a pneumatic regulating valve to obtain valve rod displacement X, pressure sensors are arranged at the front and the back of a transmission pipeline to obtain valve front pressure P1 and valve back pressure P2, and a flow velocity sensor is arranged in the pipeline to obtain medium flow velocity F. The current signal input range of the valve actuator is 4-20mA current signal, the actuator input CV is given by 25% step, variable data are collected through the signal collecting device, and the collected data are transmitted to the upper computer through the microcomputer unit through serial port communication.

2. Data preprocessing and feature extraction

The pneumatic control valve has N states including 1 normal state and N-1 fault states; in the process of training a diagnostic model, grading test is carried out when CV changes from 0% to 100% and then back to 0% in 25% step change each time; the data collected in each test is a 5 x k matrix containing 5 variables, which is spliced into a 1 x 5k vector as an original sample; each state has Num samples, and the total number of samples T is Num N; for sample data x_iAnd (3) normalization processing, wherein the normalization method comprises the following steps:

x _ min and x _ max represent minimum and maximum values in the same type of sample.

Adding labels corresponding to states from 1 to N to the converted data, and respectively representing N different states of the regulating valve; dividing the sample set into m training sample sets and n testing sample sets, wherein m + n is T; respectively carrying out one-hot coding on the training sample label and the test sample label to generate an output matrix T_m、T_n。

Standardizing the training sample and the test sample to ensure that the mean value of the samples is 0 and the variance is 1; initializing a random unit vector omega; denote training samples by X, X₂Representing a test sample, considering the dynamic relation of valve process data, and considering that the valve data is composed of dynamic characteristics and static characteristics; therefore, firstly, extracting dynamic characteristics of a data set by using a dynamic internal principal component analysis method, extracting expected number of principal elements from training set data and recording a load vector omega to form a matrix, wherein the extracted rest part X is regarded as a static component of the data, the dimension of the X is reduced to 95% of sample information by using the principal component analysis method, and the two parts of data are combined into a training set characteristic matrix; extracting the same number of principal elements from the test set according to a load matrix formed by the load vectors omega of the training set, and obtaining the rest part X₂The test set is considered as a static component, the dimension is reduced to 95% of sample information by a principal component analysis method, and the two parts of data are combined into a test set characteristic matrix.

The method for extracting the characteristics of the training sample and the test sample comprises the following steps:

2.1, extracting X principal component r of sample_i＝X_iω, wherein X_i＝[x_ix_i+1…x_N+i-1]I ═ 1,2, …, s +1, record the load vector ω;

2.2 updating autoregressive model parameter β ═ r₁,r₂,…,r_s]^Tr_s+1(ii) a Wherein T represents a transposition operation;

2.3 updating load vectors

Unitized load vector

Unitized autoregressive model parameters

2.4 scaling samples

Wherein

2.5, returning to 2.1, and extracting the next principal element until the desired number of principal elements are extracted;

2.6, regarding the rest X as a static part of the data, reducing the dimension from PCA to contain 95% of sample information;

2.7, combining the extracted principal component and the sample subjected to PCA dimensionality reduction into a training set characteristic matrix;

2.8 extracting test sample X₂Principal component r_i＝X_2iω，X_2iComposition and X_iSimilarly, where ω is consistent with the 2.1 record;

2.9, zoom sample

Wherein

2.10, returning to 2.8, and extracting the next principal element until the expected number of principal elements are extracted;

2.11, for the remaining X₂Considered as the static part of the data, reduced from PCA to contain 95% sample information;

and 2.12, combining the extracted principal component and the sample subjected to PCA dimension reduction into a test set characteristic matrix.

3. Establishing a low-deviation random configuration network fault diagnosis model:

and 3.1, initializing parameters. Maximum number of hidden layer nodes L_maxCritical tolerance error epsilon, gamma epsilon (0,1) and maximum configuration time T_maxRange of random parameters [ - λ, λ [ - ]]Changing step length delta lambda;

3.2, randomly configuring input weight w and bias b by using a low-deviation sobol sequence; building a sobol candidate point pool with the size of T_maxAccording to the formula w ═ sobol1 × 2 λ - λ, b ═ sobol2 × 2 λ - λ, T is randomly arranged in a low-deviation sequence_maxThe input weight w and the bias b of the secondary hidden layer form a candidate node pool, the low deviation point successfully forming the candidate node is removed from the sobol candidate point pool, and meanwhile, a new unused low deviation point is inserted into the sobol candidate point pool, so that the size of the sobol candidate point pool is kept unchanged; the low-deviation sobol sequence is generated as follows: set sobol sequence by

Composition v_iI.e. the i-th number, m, of sobol sequences_i＜2ⁱAnd is positive odd; v. of_i，m_iDepends on a simple polynomial f (x) of coefficient a, order p, with coefficients of only 0 or 1, of the form f (x) ═ x^p+a₁x^p-1+…+a_p-1x+a_p(ii) a For i > p, v_i、m_iThe recurrence formula is:

wherein

Meaning that the rounding is done down,

representing binary bitwise XOR to obtain a sequence which is a sobol sequence;

3.3 calculating output matrix h_LSelecting Sigmoid as the activation function, inputting the Sigmoid into the training set characteristic matrix in the step two, and calculating xi according to a supervision mechanism_LRecord T_maxSecond inner xi_LMaximum of w, b, h_L(ii) a If T_maxIf the supervision mechanism is not satisfied, the parameter γ ═ γ + τ is updated, where τ ∈ (0,1- γ), λ_newλ + Δ λ, random parameter variation is [ - λ_new，λ_new]Returning to the previous step; xi_LThe calculation formula is as follows:

wherein mu_L＝(1-γ)/(L+1)

Computing hidden layer output weights

Wherein H_L＝[h₁,h₂,…,h_L]，

Representing a generalized inverse. Residual error is e_L＝H_L·β_L-T_mWhen the residual error meets the tolerance error epsilon or the number of hidden nodes reaches L_maxWhen the node is not newly added, the modeling is completed, and the low-deviation random configuration network model f ═ H beta is obtained_L(ii) a Otherwise, returning to the previous step to continue constructing the model.

Has the advantages that: the pneumatic control valve fault diagnosis device is simple in structure and obvious in effect, the fault diagnosis accuracy is improved on the basis of realizing the fault self-diagnosis of the pneumatic control valve, the fault diagnosis missing rate is particularly reduced, and the fault operation of the valve is avoided; the universality is good, and the fault diagnosis can be completed without complex expert experience knowledge storage; general operating personnel can master, has improved pneumatic governing valve fault diagnosis's automation and intelligent degree.

Drawings

FIG. 1 is a flow chart of the present invention;

FIG. 2 is a pneumatic regulator valve for use with the present invention;

FIG. 3 shows the diagnosis results of the present invention;

in the figure: the valve comprises a spring, an air chamber, a controller, a valve rod and a valve seat.

Detailed Description

The present invention will be described in detail below with reference to the accompanying drawings.

As shown in fig. 1, the method for diagnosing the fault of the pneumatic control valve of the low-deviation random configuration network provided by the invention comprises the following steps:

1. data acquisition

For the pneumatic regulating valve shown in fig. 2, a displacement sensor is arranged at the reserved position of the regulating valve to obtain the valve rod displacement X, pressure sensors are arranged at the front and the rear of a transmission pipeline to obtain the pressure P1 before the valve and the pressure P2 after the valve, and a flow rate sensor is arranged in the pipeline to obtain the medium flow rate F. The input range of the current signal of the valve actuator is 4-20mA current signal, the input CV of the actuator is given by 25% steps, the input of the actuator is 0%, 25%, 50%, 75%, 100%, 75%, 50%, 25% and 0% in sequence, the data of each sensor is stably acquired for 30 times after each step, the data of each variable is acquired through a signal acquisition device, and the acquired data is transmitted to an upper computer through a microcomputer unit through serial port communication.

2. Data preprocessing and feature extraction

In the embodiment, the pneumatic regulating valve has 13 states, including 1 normal state and 12 fault states; in the process of training a diagnostic model, the step change of the CV is 25 percent each time, and the step change is from 0 percent to 100 percent and then back to 0 percent to be a grading test; each time of grading testArranging variable data into an original sample in sequence; 300 samples are collected in each state, and different samples in the same state can be collected through different fault intensities; each state has 300 samples, and the total number of samples T is 300 × 13; for sample data x_iAnd (3) normalization processing, wherein the normalization method comprises the following steps:

x _ min and x _ max represent minimum and maximum values in the same type of sample.

The collected data is a 5 x 270 matrix containing 5 variables, and the matrix is spliced into a 1 x 1350 vector according to rows to serve as a training sample; adding labels corresponding to states from 1 to 13 to the converted data, wherein the labels represent N different states of the regulating valve respectively; dividing the sample set into 2600 training sample sets and 1300 testing sample sets; respectively carrying out one-hot coding on the training sample label and the test sample label to generate an output matrix T_m、T_n。

Standardizing the training sample and the test sample to ensure that the mean value of the samples is 0 and the variance is 1; initializing a random unit vector omega; denote training samples by X, X₂Representing the test specimen, the following steps are completed:

2.1, extracting X principal component r of sample_i＝X_iω, wherein X_i＝[x_ix_i+1…x_N+i-1]I ═ 1,2, …, s +1, record the load vector ω;

2.2 updating autoregressive model parameter β ═ r₁,r₂,…,r_s]^Tr_s+1(ii) a Wherein T represents a transposition operation;

2.3 updating load vectors

Unitized load vector

Unitized autoregressive model parameters

2.4 scaling samples

Wherein

2.5, returning to 2.1, and extracting the next principal element until the desired number of principal elements are extracted;

2.6, regarding the rest X as a static part of the data, reducing the dimension from PCA to contain 95% of sample information;

2.7, combining the extracted principal component and the sample subjected to PCA dimensionality reduction into a training set characteristic matrix;

2.8 extracting test sample X₂Principal component r_i＝X_2iω，X_2iComposition and X_iSimilarly, where ω is consistent with the 2.1 record;

2.9, zoom sample

Wherein

2.10, returning to 2.8, and extracting the next principal element until the expected number of principal elements are extracted;

2.11, for the remaining X₂Considered as the static part of the data, reduced from PCA to contain 95% sample information;

and 2.12, combining the extracted principal component and the sample subjected to PCA dimension reduction into a test set characteristic matrix.

3. Training a low-deviation random configuration network model:

3.1, initialization parameters: setting the maximum hidden layer node number 150, the critical tolerance error epsilon is 0.15, gamma is 0.999, and the maximum configuration time T_max600, random parameter range [ -1, 1 []The change step length delta lambda is 10;

3.2 Low-bias sobol sequence random configuration input weight w and biasB, placing; building a sobol candidate point pool with the size of T_maxAccording to the formula w ═ sobol1 × 2 λ - λ, b ═ sobol2 × 2 λ - λ, T is randomly arranged in a low-deviation sequence_maxThe input weight w and the bias b of the secondary hidden layer form a candidate node pool, the low deviation point successfully forming the candidate node is removed from the sobol candidate point pool, and meanwhile, a new unused low deviation point is inserted into the sobol candidate point pool, so that the size of the sobol candidate point pool is kept unchanged; the low-deviation sobol sequence is generated as follows: set sobol sequence by

Composition v_iI.e. the i-th number, m, of sobol sequences_i＜2ⁱAnd is positive odd; v. of_i，m_iDepends on a simple polynomial f (x) of coefficient a, order p, with coefficients of only 0 or 1, of the form f (x) ═ x^p+a₁x^p-1+…+a_p-1x+a_p(ii) a For i > p, v_i、m_iThe recurrence formula is:

wherein

Meaning that the rounding is done down,

representing binary bitwise XOR to obtain a sequence which is a sobol sequence;

3.3 calculating output matrix h_LSelecting Sigmoid as the activation function, inputting the Sigmoid into the training set characteristic matrix in the step two, and calculating xi according to a supervision mechanism_LRecord T_maxSecond inner xi_LMaximum of w, b, h_L(ii) a If T_maxIf none of the times satisfies the supervision mechanism, thenUpdating the parameter γ ═ γ + τ where τ ∈ (0,1- γ), λ ∈_newλ + Δ λ, random parameter variation is [ - λ_new，λ_new]Returning to the previous step; xi_LThe calculation formula is as follows:

wherein mu_L＝(1-γ)/(L+1)

Computing hidden layer output weights

Wherein H_L＝[h₁,h₂,…,h_L]，

Representing a generalized inverse. Residual error is e_L＝H_L·β_L-T_mWhen the residual error meets the tolerance error epsilon or the number of hidden nodes reaches L_maxWhen the node is not newly added, the modeling is completed, and the low-deviation random configuration network model f ═ H beta is obtained_L(ii) a Otherwise, returning to the previous step to continue constructing the model.

4. Using the network model to diagnose the test set sample:

and calculating hidden layer output weight H. And selecting a Sigmoid function as the activation function, inputting the Sigmoid function into a test set feature matrix in 2.12 to obtain a model output matrix Y-H-beta_L(ii) a Decoding the output matrix, setting the maximum element of each row as 1, and setting the rest as 0 to obtain a converted output matrix; then with the test sample label T after single hot coding_nThe accuracy of the comparison calculation is calculated according to the formula

The model diagnosis result can be visually observed through the confusion matrix, the diagnosis result confusion matrix of the embodiment is shown in figure 3, and the diagnosis accuracy reaches 90.1%.

The foregoing is only a preferred embodiment of this invention and it should be noted that modifications can be made by those skilled in the art without departing from the principle of the invention and these modifications should also be considered as the protection scope of the invention.

Claims (3)

1. A pneumatic control valve fault diagnosis method based on a low-deviation random configuration network is characterized by comprising the following steps: the method comprises the following steps:

step one, data acquisition:

collecting valve rod displacement X, valve front pressure P1, valve back pressure P2, medium flow rate F and actuator input CV data through a signal collecting device, and then transmitting the CV data to an upper computer;

step two, data preprocessing and feature extraction:

the pneumatic control valve has N states including 1 normal state and N-1 fault states; in the process of training a diagnostic model, changing CV from 0% to 100% and then back to 0% at each time is a grading test; the data collected in each test is a 5 x k matrix containing 5 variables, which is spliced into a 1 x 5k vector as an original sample; each state has Num samples, and the total number of samples T is Num N; for sample data x_iNormalization processing; adding labels corresponding to states from 1 to N to the converted data, and respectively representing N different states of the regulating valve; dividing the sample set into m training sample sets and n testing sample sets, wherein m + n is T; respectively carrying out one-hot coding on the training sample label and the test sample label to generate an output matrix T_m、T_n(ii) a Standardizing the training sample and the test sample to ensure that the mean value of the samples is 0 and the variance is 1; denote training samples by X, X₂Representing a test sample, respectively extracting X dynamic characteristics and static characteristics by using a dynamic internal principal component analysis method and a principal component analysis method to form a training set characteristic matrix, obtaining load vectors to form a load matrix, and completing X by using the obtained load matrix₂Extracting the features of the test set to obtain a test set feature matrix;

step three, establishing a low-deviation random configuration network fault diagnosis model:

and randomly distributing input weight w and bias b by a low deviation sequence, establishing a hidden node candidate node pool according to a supervision mechanism, and selecting an optimal node to access a random configuration network each time until a termination condition is reached.

2. A pneumatic valve fault diagnosis method based on low-deviation stochastic configuration network as claimed in claim 1, wherein: in the first step, a displacement sensor is arranged at a reserved position of a pneumatic regulating valve to obtain valve rod displacement X, pressure sensors are arranged at the front and the back of a transmission pipeline to obtain pre-valve pressure P1 and post-valve pressure P2, and a flow velocity sensor is arranged in the pipeline to obtain medium flow velocity F; the input range of the current signal of the valve actuator is 4-20mA current signal, the actuator is given with 25% step input CV, and the acquired data is transmitted to the upper computer through the microcomputer unit by serial port communication.

3. A pneumatic valve fault diagnosis method based on low-deviation stochastic configuration network of claim 1, wherein the third step comprises the steps of:

3.1, initialization parameters: maximum number of hidden layer nodes L_maxCritical tolerance error epsilon, gamma epsilon (0,1) and maximum configuration time T_maxRange of random parameters [ - λ, λ [ - ]]Changing step length delta lambda;

3.2, randomly configuring input weight w and bias b by using a low-deviation sobol sequence; building a sobol candidate point pool with the size of T_maxAccording to the formula w ═ sobol1 × 2 λ - λ, b ═ sobol2 × 2 λ - λ, T is randomly arranged in a low-deviation sequence_maxThe input weight w and the bias b of the secondary hidden layer form a candidate node pool, the low deviation point successfully forming the candidate node is removed from the sobol candidate point pool, and meanwhile, a new unused low deviation point is inserted into the sobol candidate point pool, so that the size of the sobol candidate point pool is kept unchanged; the low-deviation sobol sequence is generated as follows: set sobol sequence by

Composition v_iI.e. the i-th number, m, of sobol sequences_i＜2ⁱAnd is positive odd; v. of_i，m_iDepends on a simple polynomial f (x) of coefficient a, order p, with coefficients of only 0 or 1, of the form f (x) ═ x^p+a₁x^p-1+…+a_p-1x+a_p(ii) a For i > p, v_i、m_iThe recurrence formula is:

wherein

Meaning that the rounding is done down,

representing binary bitwise XOR to obtain a sequence which is a sobol sequence;

3.3 calculating output matrix h_LSelecting Sigmoid as the activation function, inputting the Sigmoid into the training set characteristic matrix in the step two, and calculating xi according to a supervision mechanism_LRecord T_maxSecond inner xi_LMaximum of w, b, h_L(ii) a If T_maxIf the supervision mechanism is not satisfied, the parameter γ ═ γ + τ is updated, where τ ∈ (0,1- γ), λ_newλ + Δ λ, random parameter variation is [ - λ_new，λ_new]Returning to the previous step; xi_LThe calculation formula is as follows:

wherein mu_L＝(1-γ)/(L+1)

Computing hidden layer output weights

Wherein H_L＝[h₁,h₂,…,h_L]，

Representing a generalized inverse. Residual error is e_L＝H_L·β_L-T_mWhen the residual error meets the tolerance error epsilon or the number of hidden nodes reaches L_maxWhen the node is not newly added, the modeling is completed, and the low-deviation random configuration network model f ═ H beta is obtained_L(ii) a Otherwise, returning to the previous step to continue constructing the model.

Images