Typical early fault detection method based on multi-subspace division

Through the application of Kullback-Leibler divergence and sparse mutual information matrix in the subspace, combined with weighted average moving window and NPE algorithm, early fault detection problems are solved, detection accuracy and conventional fault detection capabilities are improved, which is significantly better than traditional Multivariate statistical method.

CN118916667APending Publication Date: 2024-11-08EAST CHINA UNIV OF SCI & TECH
View PDF 0 Cites 2 Cited by

Patent Information

Application Number
CN202410918379.8
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-07-10
Publication Date
2024-11-08

AI Technical Summary

Technical Problem

The prior art is difficult to detect early failures, especially because of the weak characteristics of early failures, the lack of obvious changes in variables, and the detection effect of traditional multivariate statistical methods in complex industrial processes is not ideal.

Method used

Comparison of the probability distribution differences between normal data sets and early failure data sets through Kullback-Leibler divergence, select sensitive variables to build a specific subspace, and use sparse mutual information matrix to remove redundant variables in the subspace, combining the weighted average moving window and The NPE algorithm extracts data characteristics and establishes monitoring statistics for fault detection.

Benefits of technology

It improves the detection ability of typical early faults, while maintaining the detection accuracy of conventional faults, significantly improving the detection effect and reducing noise interference.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure BDA0004935858670000026
    Figure BDA0004935858670000026
  • Figure BDA0004935858670000027
    Figure BDA0004935858670000027
  • Figure BDA0004935858670000028
    Figure BDA0004935858670000028
Patent Text Reader

Abstract

The invention discloses a typical early-stage fault detection method based on multi-subspace division, and aims to divide process variables through historical working conditions of typical early-stage faults and implement synchronous detection of conventional faults and early-stage faults on the basis of the process variables. Specifically, firstly, the KL divergence between corresponding variables is calculated by adopting normal working condition data and historical data of typical early faults difficult to detect, and sensitive variables of the early faults are divided into different subspaces; removing redundant variables from all working condition data by using a sparse mutual information matrix to form a conventional fault subspace; a weighted sliding window is used in each subspace to realize data enhancement, and then data features are extracted through NPE to realize subspace fault detection; and finally establishing a comprehensive index to implement overall fault detection. Compared with a traditional fault detection method, the method has the advantages that on the premise that the conventional fault detection rate is kept stable, the detection effect on typical early faults is obviously improved, and the method is a better early fault detection method.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to a data-driven fault detection method, and in particular to a typical early fault detection method based on multi-subspace division. Background Art

[0002] Modern process industry is developing rapidly, and the requirements for its safety and reliability are gradually increasing. It is particularly important to detect and warn the faults of complex industrial systems. In recent years, data-driven multivariate statistical process monitoring has been favored by researchers. Commonly used algorithms include principal component analysis (PCA), partial least squares (PLS), and neighborhood preserving embedding (NPE). However, compared with conventional faults, early faults lack obvious characteristics and are hidden, random, and have small deviations from the normal state. However, they are difficult to detect, but the deterioration of normal working conditions gradually increases over time. At the same time, due to the number of control units and the complexity of variables in the entire industrial process, if only traditional multivariate statistical methods are used, the effect of detecting early faults is not ideal. In order to avoid directly detecting early faults from complex industrial processes and improve the accuracy of detection, multi-subspace distributed methods are an effective solution.

[0003] Although the multi-subspace distributed method has been proven to be effective, it is not performance-driven and cannot achieve optimal detection performance without considering fault information. Therefore, it is necessary to find out the variables related to typical early faults, select the best variables, and establish a specific model for each typical early fault. KL (Kullback-Leibler divergence, KLD) can compare the differences between two probability distributions and is more sensitive than existing multivariate statistics. The present invention compares the probability distribution differences between the normal data set and the typical early fault data set by KL divergence, and selects the most sensitive variables to construct the optimal subspace related to the fault. At the same time, the mutual information matrix is ​​sparsely processed, and redundant variables are removed through the correlation strength relationship between all process variables, noise interference is reduced, and it is divided into conventional subspaces, thereby ensuring the detection capability for conventional faults. Then, in the subspace, a weighted average moving window is used to realize data reconstruction enhancement, and data features are extracted by the NPE algorithm to realize fault detection. In summary, the method of the present invention not only maintains the detection accuracy for conventional faults, but also has the detection capability for typical early faults that are difficult to detect by traditional multivariate statistical methods. Summary of the invention

[0004] The main technical problems to be solved by the present invention are: first, the typical early fault features are weak and the variable changes are not obvious. The variables sensitive to the typical early faults that are difficult to detect are selected through KL divergence to construct a specific subspace, and data enhancement is performed in the subsequent subspace detection link to improve the detection capability. The second is to ensure the detection capability of conventional faults. The sparse mutual information matrix is ​​used to remove redundant variables to construct a conventional subspace. On the basis of data enhancement, the NPE algorithm is introduced to extract the manifold information of the data, and the obtained feature information is used for modeling. Finally, Bayesian reasoning fusion is used to establish the corresponding statistics and monitor.

[0005] The technical solution adopted by the present invention to solve the above problems is: a typical early fault detection method based on multi-subspace division, comprising the following steps:

[0006] (1) Collect sample data under normal operating conditions and historical operating condition data of typical early failures in industrial processes, where the normal operating process data set is A typical early failure data set is Calculate the normal data set The mean μ and standard deviation σ of , and Z-score normalization according to formula (1) are used to obtain the training data X∈R N×p , Y∈R N×p ;

[0007]

[0008] Where x and y are the data of a single sampling point in the standardized data set, that is, a row of data in the data set X and Y, where x∈R 1×p , y∈R 1×p ;

[0009] (2) KLD is used to quantify the difference in variable distribution under normal and fault conditions, and sensitive variables are selected to construct a specific subspace corresponding to typical early faults. The specific implementation process is as follows:

[0010] ① Calculate the corresponding variables x in the two data sets one by one i and i The KLD between (i=1, 2, ..., p) is usually expressed as:

[0011]

[0012] ② Considering that the basic definition of KLD is asymmetric, it is not suitable for use as a distance metric. In practical applications, an improved symmetric form is used:

[0013]

[0014] ③ The process variables with larger KLD are considered to have greater impact on faults. Variables sensitive to typical early faults of the same type will be placed in the same sub-block. i The contribution to the jth typical early fault is as follows:

[0015]

[0016] ④ Sort the variables by contribution rate from large to small, and calculate the cumulative contribution rate (CCR) as follows:

[0017]

[0018] ⑤ Take the first d sensitive variables greater than the CCR limit to construct the specific subspace B of the jth typical early fault j ;

[0019] (3) For the normal operating condition data set X, the sparse mutual information matrix is ​​used to eliminate redundant variables in the process variables, reduce noise interference, and construct the conventional fault subspace. The specific operation process is as follows:

[0020] ① Mutual information is used to measure the linear and nonlinear relationship between two variables. i , x k )(i, k = 1, 2, ..., p) to determine the strength of the correlation between variables. The mutual information matrix is ​​defined as follows:

[0021]

[0022] ② Design a threshold to reduce the mutual information less than the threshold to zero, and perform a sparse transformation on the MI matrix to obtain a new matrix MI s , the sparse matrix MI s Converted into a distance metric matrix, the greater the mutual information, the smaller the distance, and the more similar the samples are. The formula is as follows:

[0023]

[0024] ③ After obtaining the distance measurement matrix, the hierarchical clustering method is used to classify different variables according to the distance size, ensuring that the strongly correlated variables are divided into the same subspace, eliminating the interference of redundant variables, and constructing the conventional fault subspace B N1 , B N2 , ...;

[0025] (4) Use the weighted average moving window in each subspace to reconstruct the data and achieve data enhancement. The specific operation process is as follows:

[0026] ① Use weighted average moving window to realize data enhancement. For the data sample at time t, it is necessary to determine a moving window of length d and assign different weights to the data in the window. The data with faults introduced before time t-d+1 is processed as follows:

[0027] x′ t =w t x t +Δt+w t-1 x t-1 +Δ(t-1)+…+w t-d+1 x t-d+1 +Δ(t-d+1) (8)

[0028] ② Where Δt, Δ(t-1) and Δ(t-d+1) are the deviations at the corresponding moments of fault occurrence, w t is the weight of the data in the normalized window, as shown in the following formula:

[0029]

[0030] (5) The enhanced process data is extracted through NPE while maintaining the manifold structure, constructing monitoring statistics, and establishing a subspace fault detection model. The specific operation process is as follows:

[0031] ① Calculate the Euclidean distance between each data and the other n-1 data, and then select the nearest k data as neighbors. If x a is x b neighbors, then a straight line connects node a and node b;

[0032] ②W is the weight matrix to be calculated, the weight between node a and node b is W ab , when there is no straight line connection between two different nodes, the weight value is set to zero. If there is a straight line connection between two different nodes, the weight value W is obtained by minimizing the following objective function ab :

[0033] min∑ a ||x(a)-∑ b W ab x(b)||=∑ a ||∑ b W ab (x(a)-x(b))|| (10)

[0034] Among them, ∑ b W ab =1, j=1, 2, ..., n(W ab =0(a=b)) (11)

[0035] ③ The projection vector p is obtained by calculating the eigenvalue and eigenvector of the following formula, where X E This is the subspace dataset after data enhancement:

[0036]

[0037] Where, M = (I-W) T (IW)(13)

[0038] ④Therefore, the eigenvalue corresponding to the eigenvalue set P = [p1, p2, ..., p n ], then, the monitoring statistics in the subspace are established as:

[0039] T 2 =x T PP T x (14)

[0040] (6) Given a confidence level α, take 95% to determine the control limits of each fault subspace

[0041] Steps (1) to (6) are the offline modeling stage of the method of the present invention, and steps (7) to (11) shown below are the online monitoring process of the method of the present invention;

[0042] (7) Get online samples Using the mean and standard deviation of the normal data set to standardize it, we get:

[0043] (8) According to the subspace partitioning criteria obtained in the offline modeling steps (2) and (3), the online data sample x new Divide to obtain sample data in each subspace;

[0044] (9) In each subspace, first perform data enhancement on the input data according to step (4), and calculate the statistics of the online samples of each subspace according to the projection matrix obtained in the offline modeling stage (5) after the enhanced data. The specific formula is as follows:

[0045]

[0046] (10) For the conventional fault subspace B N1 , B N2 , ..., after obtaining the statistical data of each subspace in step (9) Afterwards, each subspace needs to statistically merge the statistical data through Bayesian inference fusion to construct a comprehensive monitoring statistic:

[0047] ①

[0048] in,

[0049] P(X b )=P(X b |N)P(N)+P(X b |F)P(F) (18)

[0050] ②Conditional probability P(X b |F) is defined as follows:

[0051]

[0052] J b,new (b=N1,N2,......Nn) is the statistical data of the conventional fault subspace in the online detection data set, J b,lim为 The control limits of the corresponding conventional subspace in step (6);

[0053] (11) Online monitoring is performed according to the following judgment logic: If the monitoring statistics of the typical early failure subspace of the online collected samples If the system has a typical early fault 1, it needs to deal with the related faults in time. Similarly, it can be judged whether typical early faults 2 and 3 have occurred. If BIC>confidence level α, a conventional fault has occurred in the system, otherwise no fault has occurred in the system.

[0054] Compared with the traditional multivariate statistical method, the advantages of the method of the present invention are:

[0055] Considering the sensitivity of KL divergence to small differences between different data distributions, by calculating the KL divergence of data samples of historical working conditions and normal working conditions of typical early faults, a specific early fault subspace is divided, so that the subsequent detection rate of early faults is significantly improved. At the same time, a sparse mutual information matrix is ​​used to exclude redundant variables and divide the conventional fault subspace, ensuring the detection capability for conventional faults. In addition, considering the problem of weak features of fault data, the present invention uses a weighted average moving window in the subspace to achieve data enhancement, and uses NPE to extract features while maintaining the manifold structure to achieve fault detection; BRIEF DESCRIPTION OF THE DRAWINGS

[0056] Figure 1 The present invention is a flowchart for implementing the method.

[0057] Figure 2 , 3 This is a comparison chart of TE process fault monitoring details. Figure 2 For typical early failure 3, Figure 3 For conventional early faults 18, Figure (a) shows the method of the present invention, and Figure (b) shows the conventional distributed PCA-LOF fault detection method. DETAILED DESCRIPTION

[0058] The method of the present invention is described in detail below with reference to the accompanying drawings and specific implementation cases.

[0059] like Figure 1 As shown, a typical early fault detection method based on multi-subspace partitioning is shown. The following is a specific simulation platform case to illustrate the specific implementation process of the method of the present invention and its superiority over existing methods.

[0060] The application object is from the Tennessee-Eastman (TE) chemical process in the United States, and the prototype is an actual process flow of the Eastman Chemical production workshop. At present, due to the complexity of the TE process, it has been widely used as a standard experimental platform for fault detection and process monitoring research. The entire TE process includes 22 measurement variables, 12 operating variables, and 19 component measurement variables. The TE process object can simulate a variety of different fault types, such as step changes in material inlet temperature, cooling water fault changes, etc. It is generally believed that typical early faults in the TE process that are difficult to detect include faults 3, 9, and 15 as shown in Table 1, and conventional multivariate statistical methods have poor detection effects on these three. In order to monitor the process, 33 process variables as shown in Table 2 are selected, and the specific implementation steps of the present invention are described in detail in conjunction with the TE process.

[0061] Table 1: Typical early failures in the TE process.

[0062]

[0063] Table 2: TE process monitoring variables.

[0064]

[0065] The training set samples and test set samples were obtained under 48h running simulation. The fault was introduced at 8h. A total of 960 observations were collected, of which the first 160 observations were normal data. First, offline training was performed using the collected training set samples to establish a fault detection model, including the following steps:

[0066] (1) Collect sample data under normal operating conditions in industrial processes and historical operating condition data of typical early failures to form a normal sample data set Typical early failure 3 data set is Typical early failure 9 data set is Typical early failure 15 data sets are Standardize the four to get X∈R 960 ×33 ,Y1∈R960×33 ,Y2∈R 960×33 ,Y3∈R 960×33 ;

[0067] (2) Based on the standardized training data sets X and Y, KLD is used to quantify the difference in variable distribution under normal and fault conditions, and sensitive variables are selected to construct specific subspaces corresponding to three typical early faults, as shown in Table 3:

[0068] Table 3. TE process variables included in the three specific subspaces

[0069]

[0070] (3) For the normal operating condition data set X, the sparse mutual information matrix is ​​used to extract the correlation of process variables. After being converted into a distance metric matrix, the conventional fault subspace is divided using hierarchical clustering. The division results are shown in Table 4:

[0071] Table 4. TE process variables included in the conventional fault subspace

[0072]

[0073] (4) In the six divided subspaces, the weighted average moving window is used to enhance the data of the divided process variable data, and the enhanced data X of each subspace is obtained. n,E [n∈(1,2,3,N1,N2,N3)];

[0074] (5) In each subspace, the enhanced process data is extracted through NPE to maintain the manifold structure and construct the subspace monitoring statistic T 2 , establish a fault detection model;

[0075] (6) Given a confidence level α, take 95% to determine the control limits of each fault subspace

[0076] (7) For samples obtained online Use the mean and standard deviation of the training data set to standardize it and get x new ;

[0077] (8) According to the construction method of steps (1)-(5) of the offline process, the fault detection model parameters in the same subspace are used to calculate the online monitoring statistics

[0078] (9) For the conventional fault detection link, after obtaining the statistical data of the three conventional fault subspaces, the statistical data are statistically merged through Bayesian inference fusion to construct the comprehensive monitoring statistic BIC;

[0079] (10) Online monitoring is performed according to the following judgment logic: If the monitoring statistics of the typical early failure subspace of the online collected samples If the system has a typical early fault 1, it needs to deal with the related faults in time. Similarly, it can be judged whether typical early faults 2 and 3 have occurred. If BIC> confidence level α, a conventional fault has occurred in the system, otherwise no fault has occurred in the system.

[0080] This method is compared with the conventional distributed PCA-LOF fault detection method. The detection results of the two methods for typical early fault 3 in TE are shown in Table 5;

[0081] Table 5. Detection results of typical early fault 3

[0082]

[0083] The detection results of the two methods for conventional early fault 18 are shown in Table 6:

[0084] Table 6. Detection results of conventional early fault 18

[0085]

[0086] Finally, the detailed comparison diagram of the method of the present invention and the conventional distributed PCA-LOF fault detection method for typical early fault 3 and conventional early fault 18 is shown in Figure 2 , Figure 3 It can be found that the detection effect of the method of the present invention on typical early faults and conventional faults is superior to that of the conventional distributed multivariate statistical method, and the false alarm rate before the fault occurs and the missed alarm rate after the fault occurs are significantly lower than those of the control group algorithm.

[0087] The above-mentioned implementation cases are only used to explain the specific implementation of the present invention, rather than to limit the present invention. Therefore, all changes made according to the shape and principle of the present invention should be included in the scope of the present invention.

Claims

1. A typical early fault detection method based on multi-subspace partitioning, Features: The method comprises the following steps: The implementation process of the offline modeling phase is as follows: Step (1) collects sample data under normal operating conditions and historical operating condition data of typical early failures in the industrial process, where the normal operating process data set is A typical early failure data set is Where N represents the number of samples, p represents the number of variables, and R represents the dimension; Step (2) calculates the mean μ and standard deviation σ of the normal data set, and performs Z-score standardization according to formula (1) to obtain the training data X∈R N×p , Y∈R N×p ; Where x and y are the data of a single sampling point in the standardized data set, that is, a row of data in the data set X and Y, where x∈R 1×p , y∈R 1×p , o represents the number of variables, and R represents the dimension; Step (3) Use KLD to quantify the difference in variable distribution under normal and fault conditions, and select sensitive variables to construct a specific subspace B corresponding to typical early fault j j , until all the difficult-to-detect early fault subspaces are constructed; Step (4) For the normal operating condition data set X, use the sparse mutual information matrix to eliminate redundant variables in the process variables, reduce noise interference, and construct the conventional fault subspace B N1 ,B N2 ,…,B Nn ;, Step (5) Use the weighted average moving window to reconstruct data in each subspace to achieve data enhancement and obtain the subspace data enhancement matrix Step (6) will achieve enhanced process data By maintaining the manifold structure through NPE, features are extracted, monitoring statistics are constructed, subspace fault detection models and statistical data T are established. 2 ; Step (7) Given a confidence level α, take 95% and determine the control limits of each fault subspace Step (8) Obtain online samples Using the mean and standard deviation of the normal data set to standardize it, we get: Step (9) is to divide the online data sample x into the subspace according to the subspace partitioning criteria obtained in the offline modeling steps (3) and (4). new Divide to obtain sample data in each subspace; Step (10) first performs data enhancement on the input data in each subspace according to step (5), and calculates the statistics of the online samples in each subspace according to the projection matrix obtained in step (6) of the offline stage. Step (11) for the conventional fault subspace B N1 ,B N2 ,…, after obtaining the statistical data of each subspace in step (10) Afterwards, each subspace needs to statistically merge the statistical data through Bayesian inference fusion to construct a comprehensive monitoring statistic BIC; Step (12) performs online monitoring according to the following judgment logic: If the monitoring statistic of the typical early failure subspace of the online collected sample is A typical early fault 1 has occurred in the system, and related faults need to be handled in a timely manner. Similarly, it can be determined whether typical early faults 2 and 3 have occurred. If BIC>confidence level α, a conventional fault has occurred in the system, otherwise no fault has occurred in the system.

2. A typical early fault detection method based on multi-subspace partitioning according to claim 1, It is characterized in that In step (3), the difference in variable distribution between normal and fault conditions of the training data set is quantified according to the KL divergence, and sensitive variables are selected to construct a specific subspace corresponding to typical early faults. The specific implementation process is as follows: Step (3.1): Considering that the basic definition of KLD is asymmetric, an improved symmetric form is adopted to calculate the corresponding variables x in the two data sets one by one. i and i KLD between (i=1,2,……,p): Step (3.2): The process variable with larger KLD is considered to have a greater impact on the fault. Variables sensitive to the same type of typical early faults will be placed in the same sub-block. i The contribution to the jth typical early fault is as follows: Step (3.3): Sort the variables by contribution rate from large to small, calculate the cumulative contribution rate (CCR) according to formula (4), and take the first d sensitive variables greater than the CCR limit to construct the specific subspace B of the jth typical early fault j :

3. A typical early fault detection method based on multi-subspace partitioning according to claim 1, It is characterized in that In step (4), for the normal operating condition data set X, the sparse mutual information matrix is ​​used to eliminate redundant variables in the process variables, reduce noise interference, and construct a conventional fault subspace. The specific operation process is as follows: Step (4.1): Mutual information is used to measure the linear and nonlinear relationship between two variables. i ,x k )(i,k=1,2,……,p) to judge the strength of the correlation between variables. The mutual information matrix is ​​defined as follows: Step (4.2): Design a threshold value to reduce the mutual information less than the threshold value to zero, and perform a sparse transformation on the MI matrix to obtain a new matrix MI S , the sparse matrix MI S Converted into a distance metric matrix, the greater the mutual information, the smaller the distance, and the more similar the samples are. The formula is as follows: Step (4.3): After obtaining the distance measurement matrix, a hierarchical clustering method is used to classify different variables according to the distance size, ensuring that strongly correlated variables are divided into the same subspace, eliminating the interference of redundant variables, and constructing the conventional fault subspace BN 1 ,BN 2 ,…,BN n .

4. A typical early fault detection method based on multi-subspace partitioning according to claim 1, It is characterized in that In step (5), the weighted average moving window is used to reconstruct data, and the specific operation process of realizing data enhancement is as follows: Step (5.1): Use a weighted average moving window to achieve data enhancement. For the data sample at time t, it is necessary to determine a moving window of length d and assign different weights to the data in the window. The data with faults introduced before time t-d+1 is processed as follows: x′ t =w t x t +Δt+w t-1 x t-1 +Δ(t-1)+…+w t-d+1 x t-d+1 +Δ(t-d+1) (7) Step (5.2): where Δt, Δ(t-1) and Δ(t-d+1) are the deviations at the corresponding moments of the fault, respectively, and w t is the weight of the data in the normalized window, as shown in the following formula:

Citation Information

Cited By

  • Early fault detection method for incomplete complex industrial system data

    CN121614911A

  • Early fault detection method for incomplete complex industrial system data

    CN121614911B