A method for identifying fault variables in industrial processes based on support vector description

Abstract

The present invention discloses a method for identifying industrial process fault variables based on support vector description, comprising the following steps: step 1: data collection; step 2: data preprocessing; step 3: offline training of support vector data description model; step 4: calculation of the center of the hypersphere of the support vector data description model; step 5: online data collection; step 6: calculation of the offset of the support vector center of each dimension; step 7: determination of fault-related variables. The beneficial effect is that the method aims at the problem of nonlinear variable identification in industrial processes and the overall structural deviation of nonlinear variables, and proposes a fault identification method based on support vector theory. Compared with traditional methods, the fault variable identification rate is improved, and it can effectively assist professionals to locate the fault source.

Description

Technical Field

[0001] The present invention belongs to the field of industrial process fault variable identification, and in particular relates to an industrial process fault variable identification method based on support vector description. Background Art

[0002] When a system detects a fault, it is necessary to find the variables related to the fault. Especially when the fault is a nonlinear non-Gaussian process, the identification of fault variables is also an unsolved open problem. An ideal fault identification model should ensure that the identified variables contain fault variables. With this important information, it is easier to find the root cause of the fault. At present, most of the multivariate statistical process monitoring methods are based on contribution diagrams and reconstruction-based methods, focusing on the variance information of variables. However, in most nonlinear transformation processes, these methods cannot give the corresponding relationship between the original measured variables and the monitored variables, so it is difficult to solve the problem of nonlinear fault variable identification. This method aims at the overall structural deviation of nonlinear variables and proposes a fault identification method based on support vector theory to solve the fault identification problem related to industrial processes. Summary of the invention

[0003] The purpose of the present invention is to provide an industrial process fault variable identification method based on support vector description, which can solve the fault identification problem related to the industrial process by targeting the overall structural deviation of nonlinear variables.

[0004] The technical solution of the present invention is as follows: A method for identifying industrial process fault variables based on support vector description comprises the following steps:

[0005] Step 1: Data collection;

[0006] Step 2: Data preprocessing;

[0007] Step 3: Offline training of support vector data description model;

[0008] Step 4: Calculate the support vector data to describe the center of the model hypersphere;

[0009] Step 5: Online data collection;

[0010] Step 6: Calculate the offset of the support vector center of each dimension;

[0011] Step 7: Identify fault-related variables.

[0012] The step 1 includes collecting monitoring data of the industrial process under normal operation on the nuclear source platform, performing data cleaning, structuring the data, and obtaining a matrix X = [x1, x2, x3 ... x n ]∈R m×n

[0013] Among them, m represents the number of variables, n represents the number of sampling points, and R represents the real number domain.

[0014] The step 2 includes

[0015]

[0016] Among them, x i ∈X, represents the i-th normal sampling data, y i is the i-th data obtained after preprocessing, x max , x min They represent the maximum and minimum values ​​in the original samples respectively.

[0017] The step 3 includes

[0018] The normal sample is mapped into a hypersphere in the high-dimensional sample space through nonlinear mapping according to the support vector data description algorithm. The α of the training data is obtained by solving the following formula: i Values ​​and support vectors x s ,

[0019]

[0020]

[0021] st represents the constraint condition, α i represents the Lagrange multiplier of the ith constraint, K(x i , x j ) represents the inner product of two samples in the high-dimensional kernel space. By solving its quadratic programming problem, the optimal solution is According to the Karush-Kuhn-Tucker condition, 0≤α i The samples corresponding to ≤C are the support vectors of the hypersphere surface.

[0022] The step 4 includes obtaining the α of the training data to be solved in the training model process. i Values ​​and support vectors x s , variable C normal The center of is as follows:

[0023]

[0024] The support vector data description model uses Gaussian kernel function for nonlinear mapping. After Gaussian kernel mapping, the mapped data set is infinite-dimensional.

[0025]

[0026] According to Taylor expansion

[0027]

[0028] σ is a constant, so we omit it and put the above formula into Taylor expansion to get

[0029]

[0030] Among them, the nonlinear mapping of the Gaussian kernel function can be obtained

[0031]

[0032] The first dimension of the Gaussian kernel function mapping is used to approximate the nonlinear mapping of the Gaussian kernel function.

[0033] The step 5 includes collecting real-time running data on the nuclear source platform and cleaning it in real time according to steps 1 and 2.

[0034] The step 6 includes

[0035] Calculate the offset of the center of each dimension of the hypersphere CO(i) based on the center difference between the model in normal offline training and the online model

[0036]

[0037] Where c(i) normal is the value of the support vector center in the i-th dimension of normal offline training, c(i) faulty It is the value of the center of the support vector trained after the fault occurs in the i-th dimension. CO(i) is used as the offset of variable i compared with the original center variable, and all offset contributions are arranged from large to small, where CO(i)∈{1, 2…m} indicates that variable i has the largest i-th offset value. When the variable offset values ​​are added from large to small, when the total value is greater than 0.3, it is considered to be the core fault variable.

[0038] The step 7 includes determining the variables related to the fault according to the center offset of the support vector.

[0039] The beneficial effects of the present invention are as follows: This method aims at the problem of nonlinear variable identification existing in industrial processes and the overall structural deviation of nonlinear variables, and proposes a fault identification method based on support vector theory. Compared with traditional methods, the fault variable identification rate is improved, and it can effectively assist professionals in locating the fault source. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] Figure 1 This is a flow chart of an industrial process fault variable identification method based on support vector description provided by the present invention. DETAILED DESCRIPTION

[0041] The present invention is further described in detail below with reference to the accompanying drawings and specific embodiments.

[0042] The present invention provides an industrial process fault variable identification method based on support vector description, which uses big data analysis technology to assist the safe operation of industrial processes. By effectively processing historical data and online process data, the system status is accurately identified, and the accuracy and real-time performance of fault diagnosis are improved to ensure the safe operation of industrial processes.

[0043] like Figure 1 As shown, a method for identifying industrial process fault variables based on support vector description includes the following steps:

[0044] Step 1: Data Collection

[0045] The monitoring data of the industrial process under normal operation is collected on the nuclear source platform, and the data is cleaned and structured to obtain the matrix X = [x1, x2, x3…x n ]∈R m×n

[0046] Among them, m represents the number of variables, n represents the number of sampling points, and R represents the real number domain.

[0047] Step 2: Data Preprocessing

[0048]

[0049] Among them, x i ∈X, represents the i-th normal sampling data, y i is the i-th data obtained after preprocessing, x max , x min They represent the maximum and minimum values ​​in the original sample respectively. This is to convert data of different dimensions to the same dimension for analysis.

[0050] Step 3: Offline training of support vector data description model

[0051] The normal sample is mapped into a hypersphere in the high-dimensional sample space through nonlinear mapping according to the support vector data description algorithm. The α of the training data is obtained by solving the following formula: i Values ​​and support vectors x s .

[0052]

[0053]

[0054] st is the abbreviation of the English word subject to, which means constraint condition. irepresents the Lagrange multiplier of the ith constraint, K(x i , x j ) represents the inner product of two samples in the high-dimensional kernel space. By solving its quadratic programming problem, the optimal solution is According to the Karush-Kuhn-Tucker condition, 0≤α i The samples corresponding to ≤C are the support vectors of the hypersphere surface.

[0055] Step 4: Calculate the support vector data to describe the center of the model hypersphere

[0056] The α of the training data obtained through the training model i Values ​​and support vectors x s . Define variable C normal The center of is as follows:

[0057]

[0058] Where c(i) is C normal The value of the center of the support vector in the i-th dimension.

[0059] The support vector data description model usually uses Gaussian kernel function for nonlinear mapping. After Gaussian kernel mapping, the mapped data set is infinite-dimensional.

[0060]

[0061] According to Taylor expansion

[0062]

[0063] Considering that σ is a constant, we omit it and put the above formula into Taylor expansion to get

[0064]

[0065] Among them, the nonlinear mapping of the Gaussian kernel function can be obtained

[0066]

[0067] In this method, the first dimension of the Gaussian kernel function mapping is used to approximate the nonlinear mapping of the Gaussian kernel function. If the support vector data description model uses other kernel functions, the design of the hyperplane center can be flexibly transformed according to the different kernel functions.

[0068] Step 5: Online data collection

[0069] Collect real-time running data on the nuclear source platform and clean it in real time according to steps 1 and 2.

[0070] Step 6: Calculate the offset of the support vector center for each dimension

[0071] As shown in the following formula, the offset CO (Center Offset) of the center of the hypersphere in each dimension is calculated based on the center difference between the model in normal offline training and the online model.

[0072]

[0073] Where c(i) normal is the value of the support vector center in the i-th dimension of normal offline training, c(i) faulty It is the value of the center of the support vector trained after the failure occurs in the i-th dimension. CO(i) is used as the offset of variable i compared with the original center variable, and all offset contributions are arranged from large to small, where CO(i)∈{1, 2…m} indicates that variable i has the largest i-th offset value.

[0074] When the variable offset values ​​are added from large to small, when the total value is greater than 0.3, it is considered to be a core fault variable.

[0075] Step 7: Identify fault-related variables

[0076] The variables related to the fault are determined according to the center offset of the support vector, among which the variable with the largest center offset is considered to be the variable that may cause the fault, and the final cause of the fault is determined by relevant professionals.

Claims

1. A method for identifying industrial process fault variables based on support vector description, characterized in that: The steps include: Step 1: Data collection; The step 1 includes collecting monitoring data of the industrial process under normal operation on the nuclear source platform, performing data cleaning, structuring the data, and obtaining a matrix X = [x1, x2, x3 ... x n ]∈R m×n Where m represents the number of variables, n represents the number of sampling points, and R represents the real number domain; Step 2: Data preprocessing; The step 2 includes Among them, x i ∈X, represents the i-th normal sampling data, y i is the i-th data obtained after preprocessing, x max , x min Respectively represent the maximum and minimum values ​​in the original samples; Step 3: Offline training of support vector data description model; The step 3 includes The normal sample is mapped into a hypersphere in the high-dimensional sample space through nonlinear mapping according to the support vector data description algorithm. By solving the following formula, we can get the α of the training data. i Values ​​and support vectors x s , st represents the constraint condition, α i represents the Lagrange multiplier of the ith constraint, K(x i , x j ) represents the inner product of two samples in the high-dimensional kernel space. By solving its quadratic programming problem, the optimal solution is According to the Karush-Kuhn-Tucker condition, 0≤α i The samples corresponding to ≤C are the support vectors of the hypersphere surface; Step 4: Calculate the support vector data to describe the center of the model hypersphere; The step 4 includes The α of the training data obtained through the training model i Values ​​and support vectors x s , variable C normal The center of is as follows: The support vector data description model uses Gaussian kernel function for nonlinear mapping. After Gaussian kernel mapping, the mapped data set is infinite-dimensional. According to Taylor expansion σ is a constant, so we omit it and put the above formula into Taylor expansion to get Among them, the nonlinear mapping of the Gaussian kernel function can be obtained Use the first dimension of the Gaussian kernel function mapping to approximate the nonlinear mapping of the Gaussian kernel function; Step 5: Online data collection; Step 6: Calculate the offset of the support vector center of each dimension; The step 6 includes According to the center difference between the model in normal offline training and the online model, the offset CO of the center of each dimension of the hypersphere is calculated. CO(i) is used as the offset of variable i compared to the original center, and all offset contributions are arranged from large to small, where CO(i)∈{1, 2…m} indicates that variable i has the largest i-th offset value; when the variable offset values ​​are added from large to small, when the total value is greater than 0.3, it is considered to be a core fault variable; Step 7: Identify fault-related variables.

2. The method for identifying industrial process fault variables based on support vector description as claimed in claim 1, characterized in that: The step 5 includes collecting real-time running data on the nuclear source platform and cleaning it in real time according to steps 1 and 2.

3. The method for identifying industrial process fault variables based on support vector description as claimed in claim 1, characterized in that: The step 7 includes determining the variables related to the fault according to the center offset of the support vector.

Images