CN109145256B - Process monitoring method based on normative variable nonlinear principal component analysis - Google Patents
Process monitoring method based on normative variable nonlinear principal component analysis Download PDFInfo
- Publication number
- CN109145256B CN109145256B CN201811349678.5A CN201811349678A CN109145256B CN 109145256 B CN109145256 B CN 109145256B CN 201811349678 A CN201811349678 A CN 201811349678A CN 109145256 B CN109145256 B CN 109145256B
- Authority
- CN
- China
- Prior art keywords
- vector
- matrix
- component analysis
- principal component
- past
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Active
Links
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F17/00—Digital computing or data processing equipment or methods, specially adapted for specific functions
- G06F17/10—Complex mathematical operations
- G06F17/16—Matrix or vector computation, e.g. matrix-matrix or matrix-vector multiplication, matrix factorization
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
- G06F18/213—Feature extraction, e.g. by transforming the feature space; Summarisation; Mappings, e.g. subspace methods
- G06F18/2135—Feature extraction, e.g. by transforming the feature space; Summarisation; Mappings, e.g. subspace methods based on approximation criteria, e.g. principal component analysis
Abstract
The invention discloses a nonlinear dynamic process monitoring method based on normalized variable nonlinear principal component analysis, which comprises the following steps: acquiring a data matrix Y, and pre-specifying a numerical value of p and a system order n; combining Hankel matrixes of past and future observation values according to a formula; calculating covariance and mutual variance matrices of past and future observations; performing singular value decomposition on the H matrix; calculating a state vector and a residual vector; projecting the state vector to a high-dimensional feature space through explicit second-order polynomial mapping; determining a front k pivot element through characteristic value decomposition in principal component analysis; finally, calculate T 2 Statistic, combination statistic Q c And its corresponding control limits. The method provided by the invention is used for monitoring the faults of three different types in the Tennessman chemical process, and simulation results show that compared with KPCA and NDPCA, the CV-NPCA method provided by the invention has higher fault detection rate and relatively lower fault false alarm rate.
Description
Technical Field
The invention relates to a nonlinear dynamic process monitoring method in the technical field of data driving, in particular to a nonlinear dynamic process monitoring method based on explicit polynomial mapping.
Background
The traditional multivariate statistical process monitoring method is limited by the assumption that linear and measured variables conform to normal distribution, such as principal component analysis and normative variable analysis, and when the method is used for monitoring a nonlinear dynamic industrial process, higher false alarm rate of faults and lower fault detection rate can be generated. Nuclear principal component analysis based on radial basis functions has found application in many non-linear industrial processes. However, infinite-dimensional non-linear mapping based on an uncertain kernel function is inefficient and redundant.
Disclosure of Invention
The purpose of the invention is as follows: the invention aims to solve the problems of uncertainty and redundancy based on a radial basis kernel function, and provides a process monitoring method based on normative variable nonlinear principal component analysis and a nonlinear dynamic process monitoring method combining the advantages of normative variable analysis and principal component analysis.
The technical scheme is as follows: the invention discloses a process monitoring method based on normative variable nonlinear principal component analysis, which comprises the following steps:
1) obtaining a data matrix Y ∈ R m×l Pre-specifying a numerical value of p and a system order n;
2) hankel rectangle Y for combining past observation values according to formula p Hankel matrix Y of future observations f ;Y p =[y p,p+1 y p,p+2 … y p,p+N ]∈R mp×N And Y f =[y f,p+1 y f,p+2 … y f,p+N ]∈R mp×N The N moment is the column number of a Hankel matrix;
3) calculating covariance of Hankel matrix of past and future observationsSum and mutual variance matrix
6) Projecting state vectors to high dimensions by explicit second-order polynomial mappingCharacteristic space G ═ G 1 ,g 2 ,…,g N ] T ;
8) Calculating T 2 Statistic, combination statistic Q c And its corresponding control limit; if it is notOr Q c >Q UCL And (α), indicating that a fault was detected.
Further, the past observation vector y in step 1) p,r And a future observation vector y f,r The measurement values of past and future p sampling moments in the data matrix Y are combined to form:
wherein r is p +1, p +2, …, p + N, y r Are measurements from a training data set.
Further, the step 6) is specifically as follows:
the state vector is mapped to the high dimensional feature space by displaying a second order as follows:
for a two-dimensional vector, the dimensionality of the mapped state vector is increased to 5; for a vector of dimension D, the dimension of the mapped state vector can be determined by the following equation:
an m-dimensional state vector x r ∈R m After the display binomial mapping, the change is g r ∈R (m(m+3)/2) ;In a high-dimensional feature space, a matrix G epsilon R consisting of the first n state vectors can be obtained (N×(m(m+3)/2) With less non-linear behavior.
Further, the step 7) is specifically as follows:
reducing the dimensionality of the mapped state vector through conventional principal component analysis, and obtaining the most important l-dimensional latent variable and D-l residual vector:
by performing eigenvalue decomposition on the matrix S, a load vector matrix V and a diagonal matrix Lambda can be obtained:
S=G T G/(N-1)=VΛV T
wherein Λ ═ diag (λ) 1 ,λ 2 ,…,λ k );
The matrix G can be decomposed as the product of the load vector and the score vector:
wherein p is i Is a load vector t i And E represents a residual matrix, and the number k of the principal elements can be obtained by accumulating the contribution rate.
Further, the method also comprises online monitoring, and comprises the following specific steps:
1) after the observed value of the 2p sampling moment is obtained, the past observed vector y is assembled pp ;
3) The monitoring statistics do not exceed their corresponding control limits, indicating that the process is operating normally; if it is usedOr Q c >Q UCL And (α), indicating that a fault was detected.
Has the advantages that: the invention carries out standardization preprocessing on the training data through the standardization variable analysis to reduce the influence of the dynamic characteristics of the data. The state vector is then projected into a high dimensional feature space using the determined second order polynomial mapping. Simultaneously considering the linear and nonlinear changes of the monitoring process, and providing a combined statistic Q c And its control limit is determined using kernel density function estimation. The method is used for monitoring three different types of faults in the chemical process of Tennessee. Simulation results show that compared with KPCA and NDPCA, the proposed CV-NPCA method has higher fault detection rate and relatively lower fault false alarm rate.
Drawings
FIG. 1 is a schematic diagram of the monitoring method of the present invention;
FIG. 2 is a simulation comparison monitoring diagram of TE process fault 2 of the present invention;
FIG. 3 is a simulation versus monitoring diagram of the TE process fault 10 of the present invention;
figure 4 is a simulation versus monitoring graph of a TE process fault 19 of the present invention.
Detailed Description
The technical solution of the present invention will be further described in detail with reference to the following specific examples.
FIG. 1 is a schematic diagram of the monitoring method of the present invention. The proposed nonlinear dynamic process monitoring method mainly comprises three stages. Stage one, reducing the influence of data dynamic characteristics by using a CVA; step two, mapping the state vector to a high-dimensional feature space through display polynomial mapping; determining front k principal components and residual error by using PCA, and calculating T 2 And Q c Statistics are obtained.
1) Dynamic data pre-processing
The canonical variable analysis is a linear dimension reduction method based on multivariate statistical analysis, and past observation vector y p,r And a future observation vector y f,r From past and future p samples in the data matrix YThe measured values of the scales are combined to form:
wherein, y r Are measurements from a training data set.
Hankel matrix Y setting r ═ p +1, p +2, …, p + N, past and future observations p And Y f The definition is as follows:
Y p =[y p,p+1 y p,p+2 …y p,p+N ]∈R mp×N
Y f =[y f,p+1 y f,p+2 …y f,p+N ]∈R mp×N
the Hankel matrix contains N ═ l-2p +1 columns. The covariance and cross variance matrices of past and future observations may be calculated using the following equations:
the solution for the optimal linear combination can be obtained by performing singular value decomposition on the Hankel matrix H:
the state vector is a subset of the estimated canonical variables, defined as follows: state vectorWherein, V x First n columns containing V(n needs to be specified in advance).
2) State vector mapping
The state vector is mapped to the high dimensional feature space by displaying a second order as follows:
for a two-dimensional vector, the dimensionality of the mapped state vector is increased to 5; for a vector of dimension D, the dimension of the mapped state vector can be determined by the following equation:
an m-dimensional state vector x r ∈R m After the display binomial mapping, the change is g r ∈R (m(m+3)/2) 。
In the high-dimensional feature space, a matrix G epsilon R consisting of the first n state vectors can be obtained (N×(m(m+3)/2) Having less non-linear characteristics, i.e.
G=[g 1 ,g 2 ,…,g N ] T 。
3) Performing principal component analysis
And then reducing the dimensionality of the mapped state vector through conventional principal component analysis, and acquiring the most important l-dimensional latent variable and D-l residual vector.
By performing eigenvalue decomposition on the matrix S, a load vector matrix V and a diagonal matrix Lambda can be obtained:
S=G T G/(N-1)=VΛV T
wherein Λ ═ diag (λ) 1 ,λ 2 ,…,λ k )。
The matrix G can be decomposed as the product of the load vector and the score vector:
wherein p is i Is a load vector t i The number of principal elements k can be obtained by accumulating the contribution rates.
Based on the monitoring method, the specific monitoring step of the invention can be divided into two aspects of off-line training and on-line monitoring, and specifically comprises the following steps:
A. off-line training
2) hankel rectangle Y for combining past observation values according to formula p Hankel matrix Y of future observations f ;Y p =[y p,p+1 y p,p+2 …y p,p+N ]∈R mp×N And Y f =[y f,p+1 y f,p+2 …y f,p+N ]∈R mp×N The N moment is the column number of a Hankel matrix;
3) computing covariance of Hankel matrix of past and future observations Sum and mutual variance matrix
6) Projecting a state vector to a high-dimensional feature space G-G through explicit second-order polynomial mapping 1 ,g 2 ,…,g N ] T ;
8) Calculating T 2 Statistic, combination statistic Q c And its corresponding control limits. If it is notOr Q c >Q UCL And (α), indicating that a fault was detected.
B. On-line monitoring
1) After the observed value of the 2p sampling moment is obtained, the past observed vector y is assembled pp ;
3) The monitoring statistic does not exceed its corresponding control limit, indicating that the process is operating properly. If it is notOr Q c >Q UCL And (α), indicating that a fault was detected.
Simulation verification
Three different types of fault data sets are generated by utilizing a Tiannaxi Iseman chemical process simulation platform, and the effectiveness and the performance of the algorithm are verified. The Tennessee Eastman (TE) chemical process simulation platform in Tennessman of Tennessee can simulate the characteristics of nonlinearity, non-Gaussian, time-varying, multi-mode and the like of a process, and provides a standard simulation model for verifying various process modeling and control methods and fault monitoring and diagnosis methods. The TE process includes five main operating units: a reactor, a condenser, a compressor, a separator and a stripper; and also comprises 4 gas feeds, 2 main products generated by 2 gas-liquid exothermic reactions and 2 by-products generated by two derived exothermic reactions; the process mechanism is complex, and the variables are more, including multiple data fault types such as steps, random changes, slow drift, viscosity and constant positions. The process contained 41 measured variables and 12 controlled variables. Each measured variable is superimposed with additive noise to simulate the noise in an actual industrial process. Each data set has 52 variables, and the sampling time of most variables is 3 minutes; the sampling time for 14 variables was 6 minutes and for 5 variables was 15 minutes. The training and testing data set for the TE process may be downloaded from the following website: the http:// web. mit. edu/braatzgroup/TE _ process. zip data set includes 1(Fault 0) normal operation data and 20(Fault 1-Fault 20) failure operation data. Three types of test data, namely step, random variation and position fault, are adopted in the simulation example.
In order to compare the monitoring performance of the KPCA, NDPCA and CV-NPCA under the same condition, the length p of past and future observation windows of NDPCA and CV-NPCA are set to be 2; the number of principal elements for KPCA, NDPCA, and CV-NPCA are all set to 7. The width parameter c of the radial basis kernel function is chosen by reference as c 500 × D.
FIGS. 2(a) (b), 3(a) (b) and 4(a) (b) show the T generated by KPCA and NDPCA, respectively 2 And a monitoring map of the Q statistics. FIG. 2(c), FIG. 3(c) and FIG. 4(c) show the CV-NPCA generated T, respectively 2 Statistics sum Q c A monitoring graph of the combined statistics. It can be seen from the figure that NDPCA and CV-NPCA have higher sensitivity and lower perturbation in the case of fault 10. However, in the case of faults 2, 19, it is not easy to observe that the monitoring performance of that method is better.
For more accurate comparison of monitoring performance, two indexes of fault detection rate and false alarm rate are adopted in the example. The specific calculation formula is as follows:
wherein N is f Representing the number of samples exceeding the control limit; n is a radical of ab Total number of samples indicating abnormal operation state; n is a radical of no Representing the total number of samples in normal operation.
TABLE 1
TABLE 2
Statistics T for three fault types are given in tables 1 and 2, respectively 2 And Q (Q) c ) And comparing the fault detection rate with the false alarm rate. On the one hand, as can be seen from the table, the proposed CV-NPCA method has the highest statistic T in comparison with KPCA and NDPCA 2 And Q c And (4) fault detection rate. That is, the proposed CV-NPCA can account for more data variance variation in principal and residual spaces than the other two methods. On the other hand, as can be seen from Table 2, CV-NPCA achieves the lowest T of the three types of faults 2 And (4) fault false alarm rate. In the case of fault 2, KPCA and NDPCA have the same Q fault false alarm rate. The proposed CV-NPCA method achieves the lowest T only in the case of a failure of 10 2 And Q c The false alarm rate of (2). Yu's NDPCA method achieves the lowest Q false alarm rate only in the case of fault 2. Simulation result comparison shows that the proposed CV-NPCA methodNot only has relatively high fault detection sensitivity, but also has the same fault false alarm rate as KPCA and NDPCA.
Although the present invention has been described with reference to the preferred embodiments, it will be understood by those skilled in the art that various changes may be made and equivalents may be substituted for elements thereof without departing from the scope of the present invention.
Claims (3)
1. A nonlinear dynamic process monitoring method based on normalized variable nonlinear principal component analysis is characterized in that: the method comprises the following steps:
1) collecting normal operation data of Tensai chemical process of Tennessman to obtain data matrix Y ∈ R m×l Pre-specifying a numerical value of p and a system order n;
2) hankel rectangle Y for combining past observation values according to formula p Hankel matrix Y of future observations f ;
3) calculating covariance of Hankel matrix of past and future observationsSum and mutual variance matrix
6) Projecting a state vector to a high-dimensional feature space G-G through explicit second-order polynomial mapping 1 ,g 2 ,…,g N ] T (ii) a The method comprises the following specific steps:
the state vector is mapped to the high dimensional feature space by displaying a second order as follows:
for a two-dimensional vector, the dimensionality of the mapped state vector is increased to 5; for a vector of dimension D, the dimension of the mapped state vector is determined by the following equation:
an m-dimensional state vector x r ∈R m After the display binomial mapping, the change is g r ∈R (m(m+3)/2) ;In the high-dimensional feature space, obtaining a matrix G epsilon R consisting of the first n state vectors N×(m(m+3)/2) ;
7) Determining top-k pivot by eigenvalue decomposition in principal component analysisThe method comprises the following specific steps:
reducing the dimensionality of the mapped state vector through conventional principal component analysis, and acquiring an l-dimensional latent variable and a D-l residual vector:
and (3) decomposing the characteristic value of the matrix S to obtain a load vector matrix V and a diagonal matrix Lambda:
S=G T G/(N-1)=VΛV T
wherein Λ ═ diag (λ) 1 ,λ 2 ,…,λ k );
The matrix G is decomposed as the product of the load vector and the score vector:
wherein p is i Is a load vector t i Scoring vectors, wherein E represents a residual matrix, and the number k of principal elements is obtained by accumulating the contribution rate;
2. The nonlinear dynamic process monitoring method based on normative variable nonlinear principal component analysis according to claim 1, wherein the method comprises the following steps: past observation vector y in step 2) p,r And a future observation vector y f,r The measurement values of past and future p sampling moments in the data matrix Y are combined to form:
wherein r is p +1, p +2, …, p + N, y r Are measurements from a training data set.
3. The nonlinear dynamic process monitoring method based on normative variable nonlinear principal component analysis according to claim 1, wherein the method comprises the following steps: the method also comprises the following steps of online monitoring:
1) after the observed value of the 2p sampling moment is obtained, the past observed vector y is assembled pp ;
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201811349678.5A CN109145256B (en) | 2018-11-14 | 2018-11-14 | Process monitoring method based on normative variable nonlinear principal component analysis |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201811349678.5A CN109145256B (en) | 2018-11-14 | 2018-11-14 | Process monitoring method based on normative variable nonlinear principal component analysis |
Publications (2)
Publication Number | Publication Date |
---|---|
CN109145256A CN109145256A (en) | 2019-01-04 |
CN109145256B true CN109145256B (en) | 2022-09-16 |
Family
ID=64805857
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201811349678.5A Active CN109145256B (en) | 2018-11-14 | 2018-11-14 | Process monitoring method based on normative variable nonlinear principal component analysis |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN109145256B (en) |
Families Citing this family (6)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN109739214B (en) * | 2019-01-14 | 2020-08-25 | 中国石油大学(华东) | Method for detecting intermittent faults in industrial process |
CN111259523B (en) * | 2020-01-09 | 2023-03-31 | 浙江理工大学 | Process monitoring method based on KPCA-CVA model and random algorithm |
CN113377078B (en) * | 2021-04-20 | 2023-04-07 | 山东师范大学 | Industrial process fault detection method and system based on kernel principal component regression |
CN113486607B (en) * | 2021-07-01 | 2022-04-12 | 天津大学 | Gas-liquid two-phase flow state monitoring method for multi-sensor information nuclear specification variable analysis |
CN113848876B (en) * | 2021-08-13 | 2022-04-22 | 南通大学 | Low-communication and low-calculation-consumption fault detection method for multi-AGV distance cruise system |
CN115047853B (en) * | 2022-06-27 | 2023-06-06 | 中国人民解放军海军航空大学 | Minor fault detection method based on recursion standard variable residual error and kernel principal component analysis |
Citations (4)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
WO2012040916A1 (en) * | 2010-09-29 | 2012-04-05 | 东北大学 | Fault monitoring method of continuous annealing process based on recursive kernel principal component analysis |
CN107632592A (en) * | 2017-09-01 | 2018-01-26 | 南通大学 | Nonlinear time-varying procedure fault monitoring method based on efficient recursion core pivot element analysis |
CN108446529A (en) * | 2018-06-22 | 2018-08-24 | 太原理工大学 | Organic rankine cycle system fault detection method based on broad sense cross-entropy-DPCA algorithms |
CN108549908A (en) * | 2018-04-13 | 2018-09-18 | 浙江科技学院 | Chemical process fault detection method based on more sampled probability core principle component models |
-
2018
- 2018-11-14 CN CN201811349678.5A patent/CN109145256B/en active Active
Patent Citations (4)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
WO2012040916A1 (en) * | 2010-09-29 | 2012-04-05 | 东北大学 | Fault monitoring method of continuous annealing process based on recursive kernel principal component analysis |
CN107632592A (en) * | 2017-09-01 | 2018-01-26 | 南通大学 | Nonlinear time-varying procedure fault monitoring method based on efficient recursion core pivot element analysis |
CN108549908A (en) * | 2018-04-13 | 2018-09-18 | 浙江科技学院 | Chemical process fault detection method based on more sampled probability core principle component models |
CN108446529A (en) * | 2018-06-22 | 2018-08-24 | 太原理工大学 | Organic rankine cycle system fault detection method based on broad sense cross-entropy-DPCA algorithms |
Non-Patent Citations (1)
Title |
---|
基于递推规范变量分析的时变过程故障检测;商亮亮 等;《东北大学学报》;20161231;全文 * |
Also Published As
Publication number | Publication date |
---|---|
CN109145256A (en) | 2019-01-04 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN109145256B (en) | Process monitoring method based on normative variable nonlinear principal component analysis | |
Rato et al. | A systematic comparison of PCA‐based statistical process monitoring methods for high‐dimensional, time‐dependent processes | |
US7421351B2 (en) | Monitoring and fault detection in dynamic systems | |
Van den Kerkhof et al. | Dynamic model-based fault diagnosis for (bio) chemical batch processes | |
CN111222549A (en) | Unmanned aerial vehicle fault prediction method based on deep neural network | |
CN108549908B (en) | Chemical process fault detection method based on multi-sampling probability kernel principal component model | |
Wang et al. | Data-driven sensor fault diagnosis systems for linear feedback control loops | |
Ralston et al. | Computer-based monitoring and fault diagnosis: a chemical process case study | |
CN108958226B (en) | TE process fault detection method based on survival information potential-principal component analysis algorithm | |
Xie et al. | Block adaptive kernel principal component analysis for nonlinear process monitoring | |
CN111368428B (en) | Sensor precision degradation fault detection method based on monitoring second-order statistics | |
Huwang et al. | Phase II profile monitoring based on proportional odds models | |
Black et al. | System modeling and instrument calibration verification with a nonlinear state estimation technique | |
Hu et al. | Direct projection to latent variable space for fault detection | |
Lv et al. | Interpretable fault detection using projections of mutual information matrix | |
Zhao et al. | Output‐relevant fault detection and identification of chemical process based on hybrid kernel T‐PLS | |
CN116627116A (en) | Process industry fault positioning method and system and electronic equipment | |
CN112418577A (en) | Visual monitoring method and system for industrial product production process | |
Shaikh et al. | Data-driven based fault diagnosis using principal component analysis | |
Zhang et al. | A comparison of different statistics for detecting multiplicative faults in multivariate statistics-based fault detection approaches | |
CN110794814B (en) | Fault determination method and system based on generalized principal component | |
Sheikhrabori et al. | Maximum likelihood estimation of change point from stationary to nonstationary in autoregressive models using dynamic linear model | |
CN109523195B (en) | Sample multi-modal classification-based fault detection method for hierarchical non-Gaussian algorithm | |
CN110928263B (en) | Fault detection method and system for complex process considering dynamic relationship in advance | |
Salahshoor et al. | A new on-line predictive monitoring using an integrated approach adaptive filter and PCA |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
PB01 | Publication | ||
PB01 | Publication | ||
SE01 | Entry into force of request for substantive examination | ||
SE01 | Entry into force of request for substantive examination | ||
TA01 | Transfer of patent application right |
Effective date of registration: 20220119 Address after: 226000 6th floor, building 11a, Zilang science and Technology City, No. 60, Chongzhou Avenue, development zone, Nantong City, Jiangsu Province Applicant after: Baokong (Nantong) IOT Technology Co.,Ltd. Address before: 226000 No. 9 Siyuan Road, Chongchuan District, Nantong City, Jiangsu Province Applicant before: NANTONG University |
|
TA01 | Transfer of patent application right | ||
GR01 | Patent grant | ||
GR01 | Patent grant |