Hierarchical multivariable process monitoring method and device based on mutual information matrix projection

Abstract

The invention discloses a hierarchical multivariable process monitoring method and device based on mutual information matrix projection, and the method comprises the steps: generating a standard training set for modeling, then carrying out the first-layer feature projection based on a mutual information matrix, and determining the fault detection control limit of the first-layer feature projection; second-layer feature projection based on the mutual information matrix is carried out, and a principal component subspace and a residual subspace of slow feature projection are determined; determining a plurality of second control limits for dynamic monitoring of the control performance of the second-layer feature projection system; on one hand, mutual information between variables is estimated based on an alpha-entropy function of a matrix Renyi, and non-linear correlation characteristics implied in data are mined to improve the fault detection rate; and on the other hand, differential space projection is carried out on the conversion element features after mutual information matrix projection, potential mutual information slow correlation features are extracted, and the dynamic change of the system control performance can be monitored according to corresponding monitoring indexes.

Description

Technical Field

[0001] The present invention relates to the field of process monitoring, and particularly to a hierarchical multivariate process monitoring method and device based on mutual information matrix projection. Background Art

[0002] A non-ideal reactor consists of complex physical connections and control loops to form a chemical process. If a fault occurs at a certain point in the interconnected production process, it will spread and evolve along with the system interconnection. At a minimum, it will cause economic losses such as equipment downtime, production process termination, and unqualified product quality. At worst, it will trigger serious safety accidents such as explosions and cause casualties. An intelligent control scheme led by anomaly monitoring technology has played an important role in ensuring the safe operation of complex industrial processes. In particular, data-driven process monitoring technology has become one of the most active fields in process control research.

[0003] Process monitoring based on multivariate statistical analysis methods is one of the mainstream research techniques in data-driven process control, aiming to project high-dimensional observation data into a low-dimensional space based on the correlation between variables, and then reveal the fault evolution trajectory.

[0004] In the existing multivariate statistical process monitoring methods, in order to extract slow time-varying features that change over time, the process monitoring algorithm based on slow feature analysis is an unsupervised learning algorithm.

[0005] The inventors have found through research that the existing multivariate statistical process monitoring and technologies have at least the following defects:

[0006] Traditional slow feature extraction methods focus on the extraction of linear correlations, and are insufficient in mining the more essential non-linear correlation features and the temporal variation features of mutual information between variables in the dynamic process. Therefore, it is easy to have missed judgments or misjudgments when distinguishing the actual dynamic faults of the system, normal deviations under operating conditions, and the condition switching of the closed-loop controller.

[0007] The information disclosed in this background art section is only intended to increase the understanding of the overall background of the present invention, and should not be regarded as an admission or any form of suggestion that this information constitutes the prior art already known to those of ordinary skill in the art. Summary of the Invention

[0008] The purpose of the present invention is to be able to accurately distinguish the actual dynamic faults of the system, normal deviations under operating conditions, or the condition switching of the closed-loop controller, so as to improve the monitoring effect of the system control performance.

[0009] The present invention provides a hierarchical multivariate process monitoring method based on mutual information matrix projection. The steps of its offline modeling include:

[0010] S11. Perform standard preprocessing on the original data of discrete time series data to generate a standard training set for modeling;

[0011] S12. Perform the first-layer feature projection based on the mutual information matrix, including: approximately calculating the mutual information between variables for the standard training set using the α-entropy function based on matrix Rényi, constructing a mutual information matrix, and obtaining the corresponding first eigenvector matrix and first eigenvalue matrix through eigenvalue decomposition of the mutual information matrix; extracting the first transformed element features through spatial projection based on the mutual information matrix;

[0012] S13. Determine the principal component subspace and the residual subspace, including: determining the number of elements in the principal component space according to a preset eigenvalue contribution rule, denoted as the number of principal components; and obtaining the corresponding principal component eigenvector matrix, residual eigenvector matrix, principal component eigenvalue matrix, and residual eigenvalue matrix;

[0013] S14. Determine the first control limit for fault detection of the first-layer feature projection according to the monitoring statistic and confidence level of the standard training set; the first control limit is used as the fault alarm threshold;

[0014] S15. Perform the second-layer feature projection based on the mutual information matrix, including: constructing a differential space with the first transformed element features with a time delay τ; calculating the time-delay mutual information matrix on the differential space; performing eigenvalue decomposition on the time-delay mutual information matrix to obtain a second eigenvector matrix and a second eigenvalue matrix, and obtaining the first slow mutual information features based on the time-delay mutual information matrix;

[0015] S16. Determine the principal component subspace and the residual subspace of the slow mutual information feature projection, including: obtaining the principal component eigenvector matrix, residual eigenvector matrix, principal component eigenvalue matrix, and residual eigenvalue matrix of the slow mutual information feature projection according to the contribution rule of the slowness self-mutual information;

[0016] S17. Determine multiple second control limits for dynamic monitoring of the second-layer feature projection, including: calculating each of the second control limits according to the covariance matrix of the principal component slow mutual information features and the residual slow mutual information features, and the covariance matrix of the first-order differences of the principal component slow mutual information features and the residual slow mutual information features; each of the second control limits is used to judge the dynamic characteristics of the system control performance after a fault.

[0017] Preferably, in the present invention, it further includes a step of online monitoring, including:

[0018] S21. Perform standard preprocessing on the online monitoring data using the mean and variance of the original training set to generate a standard test set for model input;

[0019] S22, extracting test set conversion element features based on the mutual information matrix, including: projecting the standard test set onto a space based on the mutual information matrix to obtain second conversion element features based on the mutual information matrix;

[0020] S23, performing differential space slow mutual information feature extraction based on the mutual information matrix, including: performing differential calculation on the second conversion element feature with a time delay τ, and projecting the differential feature to the time-delay mutual information matrix space of the standard training set to obtain a second slow mutual information feature;

[0021] S24, calculating the online monitoring index, including: calculating the first monitoring index of the first layer feature projection of the standard test set, and a plurality of second monitoring indexes on the principal component subspace and the residual subspace of the second layer feature projection of the standard test set;

[0022] S25. Taking the first control limit as a threshold, judging the fault alarm according to the first monitoring index, and taking each of the second monitoring indexes as a parameter, monitoring the system control performance according to each of the second control limits, and judging the dynamic characteristics of the system after the fault.

[0023] On the other hand, a hierarchical multivariable process monitoring device based on mutual information matrix projection is provided. The hierarchical multivariable process monitoring device based on mutual information matrix projection includes a computer program stored on a medium. The computer program includes program instructions. When the program instructions are executed by a computer, the computer executes the methods described in the above aspects and achieves the same technical effects.

[0024] Compared with the prior art, the present invention has the following beneficial effects:

[0025] The present invention adopts the α-entropy function based on the matrix Rényi to estimate the mutual information between variables, realizes the real-time approximate calculation of the probability density function under process chemical industry, and performs projection based on the mutual information matrix to mine the implicit nonlinear correlation characteristics in the data, thereby obtaining a higher fault detection rate; on the other hand, the present invention performs differential space projection on the conversion element features after the mutual information projection, extracts potential slow mutual information related features, and analyzes the time series changes of the mutual information between variables. In this way, the dynamic process changes of the system can be monitored according to the corresponding monitoring indicators.

[0026] Since the hierarchical projection fault diagnosis technology based on the mutual information matrix belongs to the unsupervised process monitoring technology and the calculation using probability distribution is not affected by the uncertain factors in the chemical process, it can be widely applied to the process monitoring field of the actual chemical process. In practical applications, the present invention has achieved excellent process monitoring effects on the numerical simulation dataset and the TEP actual chemical process dataset. It can not only effectively improve the fault detection rate and reduce the false alarm rate, but also distinguish whether it is an actual dynamic fault of the system or a normal deviation under operating conditions, thus providing a reference for the subsequent processing of operators. Therefore, the present invention has important practical significance for solving the actual chemical process monitoring problem.

[0027] The above description is only an overview of the technical solution of the present invention. In order to more clearly understand the technical means of the present invention and be implemented according to the content of the specification, and to make the above and other purposes, technical features and advantages of the present invention more understandable, one or more preferred embodiments are listed below and detailed as follows in conjunction with the drawings. Brief Description of the Drawings

[0028] Figure 1 It is a step diagram of the hierarchical multivariate process monitoring method based on the projection of the mutual information matrix described in the present invention;

[0029] Figure 2 It is a fault detection result diagram based on principal component analysis for fault 14 of the TE dataset in the present invention;

[0030] Figure 3 It is a process monitoring result diagram based on slow feature analysis for fault 14 of the TE dataset in the present invention;

[0031] Figure 4 It is a hierarchical multivariate process monitoring result diagram based on the projection of the mutual information matrix for fault 14 of the TE dataset in the present invention;

[0032] Figure 5 It is a structural schematic diagram of the hierarchical multivariate process monitoring device based on the projection of the mutual information matrix described in the present invention. Detailed Description of the Preferred Embodiments

[0033] The following will describe in detail the specific embodiments of the present invention in conjunction with the drawings, but it should be understood that the protection scope of the present invention is not limited by the specific embodiments.

[0034] Unless otherwise clearly stated, throughout the specification and claims, the term "comprise" or its variations such as "comprises" or "including" etc. will be understood to include the stated elements or components, without excluding other elements or other components.

[0035] In this document, for convenience of description, spatial relative terms such as "below", "beneath", "lower", "above", "upper", etc. may be used to describe the relationship between one element or feature and another element or feature in the drawings. It should be understood that spatial relative terms are intended to encompass different directions of an object in use or operation in addition to the directions depicted in the figures. For example, if an object in the figure is flipped, an element described as "below" or "beneath" another element or feature will be oriented "above" that element or feature. Thus, the exemplary term "below" can encompass both the lower and upper directions. The object may also have other orientations (rotated 90 degrees or other orientations) and the spatial relative terms used herein should be interpreted accordingly.

[0036] In this document, terms such as "first", "second", etc. are used to distinguish two different elements or parts, and are not used to define a specific position or relative relationship. In other words, in some embodiments, the terms "first", "second", etc. may also be interchanged with each other.

[0037] Embodiment 1

[0038] In order to accurately distinguish the actual dynamic faults of the system and the normal deviations under operating conditions or the condition switching of the closed-loop controller, so as to improve the effect of monitoring the system control performance, as Figure 1 shown, in an embodiment of the present invention, a hierarchical multivariate process monitoring method based on mutual information matrix projection is provided, and the steps of offline modeling include the steps:

[0039] S11. Perform standardized preprocessing on the original data of discrete time-series data to generate a standard training set for modeling;

[0040] The specific implementation manner of this step may include:

[0041] During the operation of the chemical process, the data collected by the sensor is discrete time-series data, which is the original process data; an original training set can be constructed according to the original process data

[0042] By performing standardized processing on the original training set, a standard training set can be obtained:

[0043]

[0044] The formula for performing standardized processing includes:

[0045]

[0046] wherein, is the mean of the original training set, is the standard deviation of the original training set.

[0047] S12. Perform the first-layer feature projection based on the mutual information matrix, including: approximately calculating the mutual information between variables for the standard training set by using the α-entropy function based on matrix Rényi, constructing a mutual information matrix, and obtaining the corresponding first eigenvector matrix and first eigenvalue matrix through eigenvalue decomposition of the mutual information matrix; extracting the first transformation element features through spatial projection based on the mutual information matrix.

[0048] The specific implementation method of this step may include:

[0049] The mutual information matrix based on the α-entropy function of matrix Rényi includes:

[0050] Define the kernel function κ: Calculate the Gram matrix K = κ(x i and x j ) of the sensor variables in the standardized test set; i , x j )

[0051] Normalize the Gram matrix to obtain

[0052] Calculate the α-entropy function based on matrix Rényi according to the matrix A:

[0053]

[0054]

[0055] where represents the Hadamard product between matrices A and B;

[0056] Construct a variable mutual information matrix based on the α-entropy function of matrix Rényi, and the mutual information matrix specifically includes:

[0057]

[0058] where H(x i ) is the entropy of variable x i , H(x i ; x j ) is the joint entropy between variables x i and x j , I(x i ; x j ) is the mutual information between variables x i and x j , and I(x i ; x j ) = H(x i ) + H(xj ) - H(x i ; x j ); Specifically, on the diagonal of the mutual information matrix, there is I(x i ; x i ) = H(x i );

[0059] Perform eigen - decomposition on the mutual information matrix to obtain the first eigen - vector matrix and the first eigenvalue matrix Λ = diag(λ1, λ2, …, λ ) ∈ R m ; m×m ;

[0060] Extract the first transformation - element feature through spatial projection based on the mutual information matrix: Z train = {z train (1), z train (2), …, z train (n)},

[0061] S13. Determine the principal - component subspace and the residual subspace, including: Determine the number of elements in the principal - component space according to the preset eigenvalue contribution rule, denoted as the number of principal components; and obtain the corresponding principal - component eigen - vector matrix, residual eigen - vector matrix, principal - component eigenvalue matrix, and residual eigenvalue matrix;

[0062] The specific implementation of this step may include:

[0063] According to the eigenvalue contribution rule determine the number of elements in the principal - component space, denoted as the number of principal components I, and obtain the principal - component eigen - vector matrix W I , residual eigen - vector matrix W m-I , principal - component eigenvalue matrix Λ I and residual eigenvalue matrix Λ m-I .

[0064] S14. Determine the first control limit for fault detection of the first - layer feature projection according to the monitoring statistic and confidence level of the standard training set; the first control limit is used as the fault - alarm threshold;

[0065] The specific implementation of this step may include:

[0066] Determine the control limit for fault detection of the first - layer feature projection denoted as the first control limit: Calculate the statistical index T 2 of the standard training - set data, and obtain

[0067] T 2 = xT UΛ -1 U T x;

[0068] Among them, U represents the mutual information matrix of the eigenvector matrix, and Λ represents the corresponding eigenvalue matrix, obtained with a confidence level of α

[0069] S15. Perform the second-layer feature projection based on the mutual information matrix, including: constructing a differential space using the first transformed element feature with a time delay τ; calculating the time-delay mutual information matrix on the differential space; performing eigen-decomposition on the time-delay mutual information matrix to obtain a second eigenvector matrix and a second eigenvalue matrix, and obtaining the first slow mutual information feature based on the time-delay mutual information matrix;

[0070] The specific implementation method of this step may include:

[0071] Using the projected first transformed element feature space projection Z train ={z train (1), z train (2), …, z train (n)}, to make a differential space with a time delay τ

[0072] Calculate the time-delay mutual information matrix on the differential space

[0073]

[0074] Among them, is the mutual information between the differential space feature dimension and ;

[0075] Perform eigen-decomposition on to obtain the second eigenvector matrix and the second eigenvalue matrix

[0076] Obtain the first slow mutual information feature based on the time-delay mutual information matrix:

[0077] Y train ={y train (1), y train (2), …, y train (n)}.

[0078] S16. Determine the principal subspace and residual subspace of the slow mutual information feature projection, including: obtaining the principal eigenvector matrix, residual eigenvector matrix, principal eigenvalue matrix, and residual eigenvalue matrix of the slow mutual information feature projection according to the contribution degree rule of the slowness self-mutual information;

[0079] The specific implementation method of this step may include:

[0080] Calculate the auto mutual information AI(y 1:n-τ ; y 1+τ:n ), and screen the principal components according to the contribution degree rule of slowness auto mutual information , denoted as the number of principal components d, and obtain the principal component feature matrix W of the slow mutual information feature projection T,d , residual feature matrix W T,e , principal component eigenvalue matrix Ω d and residual eigenvalue matrix Ω e .

[0081] S17. Determine multiple second control limits for dynamic monitoring of the second layer feature projection, including: calculating each of the second control limits according to the covariance matrix of the principal component slow mutual information feature and the residual slow mutual information feature, and the covariance matrix of the first-order difference of the principal component slow mutual information feature and the residual slow mutual information feature; each of the second control limits is used to judge the dynamic characteristics of the system control performance after a fault.

[0082] The specific implementation method of this step may include:

[0083] The control limits for dynamic monitoring of the second layer feature projection are denoted as the second control limits, including and

[0084] Use and to calculate and 's statistics, use and to calculate and 's statistics; where, ∑ d and ∑ e are the covariance matrices of the slow mutual information feature y d of the principal component subspace and the slow mutual information feature y e of the residual subspace respectively; Π d and Π e are the covariance matrices of the first-order difference of the principal component slow mutual information feature and the first-order difference of the residual slow mutual information feature respectively;

[0085] Calculate the second control limits and

[0086] Furthermore, in the embodiment of the present invention, it may further include the step of online monitoring, which may specifically include:

[0087] S21. Use the mean and variance of the original training set to perform standardized preprocessing on the online monitoring data, and generate a standard test set for model input;

[0088] The specific implementation method of this step may include:

[0089] Construct a test set of the original data based on the online monitoring data in the chemical process:

[0090]

[0091] Use the mean of the original training set and variance to perform standardized preprocessing on the test set to obtain a standard test set

[0092] S22. Perform extraction of test set transformation meta-features based on the mutual information matrix, including: project the standard test set onto the space based on the mutual information matrix, and obtain the second transformation meta-feature based on the mutual information matrix: y(t) = Λ - 12 Ux(t);

[0093] S23. Perform extraction of slow mutual information features in the differential space based on the mutual information matrix, including: perform differential calculation on the second transformation meta-feature with a time delay τ, and project the differential feature onto the time-delay mutual information matrix space of the standard training set to obtain the second slow mutual information feature: y(t) = P T z(t) = P T Λ -12 Ux(t);

[0094] S24. Calculate the online monitoring indicators, including: calculate the first monitoring indicator of the projection of the first-layer features of the standard test set, and multiple second monitoring indicators on the principal component subspace and the residual subspace of the projection of the second-layer features of the standard test set, where:

[0095] The first monitoring indicator includes: T 2 = x T UΛ -1 U T x;

[0096] The second monitoring indicator includes: and where:

[0097] It should be noted that in the above formulas (i.e., steps S21 to S24), U represents the eigenvector matrix of the training set mutual information matrix Λ represents the corresponding eigenvalue matrix; yd and e Respectively represent the slow mutual information features of the principal component subspace and residual subspace of the test set, Σ d and Σ e Respectively represent y d and e The covariance matrix of and They are the first-order difference of the principal component slow mutual information feature And the first-order difference of the residual slow mutual information feature The covariance matrix of .

[0098] S25. Taking the first control limit as a threshold, judging the fault alarm according to the first monitoring index, and taking each of the second monitoring indexes as a parameter, monitoring the system control performance according to each of the second control limits, and judging the dynamic characteristics of the system after the fault.

[0099] In practical applications, fault monitoring is performed based on the monitoring value of the first layer feature projection and the first control limit. If at a certain moment T 2 The control quantity exceeds the first control limit Then a fault alarm is issued;

[0100] Next, in the embodiment of the present invention, the monitoring value index of the second layer feature projection can also be used. The process is monitored by corresponding control limits (ie, a plurality of second control limits respectively corresponding to each monitoring value indicator) to determine the dynamic characteristics of the system after the fault.

[0101] Chemical process belongs to process industry, and it is difficult to calculate the probability density function in real time by traditional methods. In the embodiment of the present invention, by using the α-entropy function based on the matrix Rényi to estimate the mutual information between variables, it provides feasibility for the real-time data mutual information calculation of process industry.

[0102] In summary, the embodiment of the present invention adopts the α-entropy function based on the matrix Rényi to estimate the mutual information between variables, realizes the real-time approximate calculation of the probability density function under process chemical industry, and performs projection based on the mutual information matrix to mine the implicit nonlinear correlation characteristics in the data, thereby obtaining a higher fault detection rate; on the other hand, the embodiment of the present invention performs differential space projection on the conversion element features after the mutual information projection, extracts potential slow mutual information related features, and analyzes the time series changes of the mutual information between variables. In this way, the dynamic process changes of the system can be monitored according to the corresponding monitoring indicators.

[0103] As can be seen from the above, the embodiment of the present invention uses the mutual information matrix projection to select the main driving quantity. The slowness mutual information based on the mutual information matrix differential space projection provides interpretability in terms of time consistency, and can further analyze the potential causes of faults. The hierarchical multivariate projection fault diagnosis technology based on the mutual information matrix in the embodiment of the present invention belongs to the unsupervised process monitoring technology, which is suitable for the characteristics of non-linearity and strong coupling in the actual process chemical process. The approximate mutual information calculation using the α-entropy function based on matrix Rényi is not affected by the uncertain factors in the chemical process, so it can be widely applied to the process monitoring field of the actual chemical process. Using the embodiment of the present invention, excellent process monitoring effects can be obtained on the numerical simulation data set and the TEP actual chemical process data set. It can not only effectively improve the fault detection rate, reduce the false alarm rate, and significantly improve the fault monitoring performance, but also monitor and distinguish the actual dynamic faults of the system, the normal deviations under operating conditions, and the condition switching of the closed-loop controller, providing reference for the operators; thus, it has important practical significance for solving the actual chemical process monitoring problems.

[0104] In a specific example of the embodiment of the present invention, the specific process and data include:

[0105] 1. Comparison between the mutual information matrix and the covariance matrix

[0106] Compared with the covariance matrix that can only capture the linear correlation between variables, the mutual information matrix M can capture the non-linear correlation between any two variables, which is just suitable for the actual multivariate chemical process. Therefore, it is of practical significance to use the mutual information matrix M projection for non-linear feature extraction.

[0107] 2. To verify the feasibility of the hierarchical multivariate process monitoring method based on the mutual information matrix projection in fault detection, a multivariate process as shown in the following formula is constructed:

[0108]

[0109] In the above formula, v represents three independent Gaussian distribution data sources with a mean of [1, 1, 3] T and a variance of [0.1, 0.1, 0.1] T e represents Gaussian white noise with a standard deviation of [0.061, 0.063, 0.198, 0.0176] T ;

[0110] where s satisfies The weight matrix β is:

[0111]

[0112] At the same time, the non-linearity between sensors is introduced,

[0113]

[0114] In the above formula, [e5, e6, e7] T represents three Gaussian distributions with a mean of 0 and a variance of 0.1; the initial value of x7 is 0.15; combining all variables gives the data set [x1, x2, …, x7] T .

[0115] Consider four types of faults that are relatively common in actual chemical processes:

[0116] Type I: Sensor degradation x * = xf, f = 0.8 k , k changes from 1.1 to 51 over time;

[0117] Type II: Sensor accuracy degradation x * = x + f, f is Gaussian white noise with a power of 1.44;

[0118] Type III: Process fault s * = s + 0.4f, f is a Gaussian distribution with a mean of 1 and a variance of 0.5;

[0119] Type IV: Dynamic change Δβ3 = [0.835, -0.820, 0.662, 0.061], where β3 represents the third row of β.

[0120] Under normal conditions, 1000 sample points are generated as the training set, and the generated test set contains 1500 sample points, and all faults are introduced at the 1001st sample point. For simplicity, sensor faults are introduced on variable x1, and process faults are introduced on data source s1.

[0121] To verify the effectiveness of the algorithm, the algorithm of this patent is compared with the principal component analysis PCA method and the slow feature analysis SFA method.

[0122] Table 1 shows the monitoring indicators of the three algorithms under four types of faults. It can be seen that the present invention maintains a high fault detection rate and a low false alarm rate under these four fault types. Specifically, for fault type III, the T 2 statistic and the SPE statistic based on the PCA method are close to failure; the statistic and the statistic based on the SFA method can successfully detect the fault, and the alarm rates are 90.38% and 93.79% respectively, but the false alarm rates are too high, 4.10% and 14.90% respectively; the monitoring quantity T of the first layer of the hierarchical diagnosis method of the embodiment of the present invention 2The statistic successfully detected the fault, with a detection rate of 96.59% and a false alarm rate of 1.80%, improving the fault detection effect. Table 1 illustrates the effectiveness of the mutual information matrix in extracting non-linear relationships compared to the covariance matrix; further, the monitoring indicators of the second layer are mainly designed to monitor the dynamic characteristics of the system after a fault occurs from different perspectives and identify changes in operating conditions, such as whether the steady state under the fault is reached.

[0123] Table 1 shows the fault detection indicators of the numerical simulation. Among them, the bolded group is the optimal group for this fault. The first and second rows of each fault correspond to the false alarm rate and the fault detection rate respectively.

[0124] Table 1:

[0125]

[0126]

[0127] 3. Performance Verification - Experimental Verification of the TE Process

[0128] The method proposed in the present invention was verified in the TE (Tennessee Eastman) process, a benchmark test set for chemical processes. The TE process contains 21 types of faults. The data used in the embodiments of the present invention can be found on the website: http: / / web.mit.edu / braatzgroup / links.html. Since faults 3, 9, 15, and 21 are minor faults and difficult to detect using statistical analysis methods, these faults will not be discussed.

[0129] Table 2 shows the fault monitoring indicators for 17 types of faults in the TE process. Compared with the PCA method and the SFA method, the embodiments of the present invention are more accurate in fault identification, as shown by the monitoring indicator T of the first-layer projection 2 which has a high detection rate and a low false detection rate for faults 1, 2, 4, 6, 7, 8, 12, 14, 17, and 18, verifying the effectiveness of the mutual information matrix in extracting non-linear relationships compared to the covariance matrix; however, the detection effect of the first layer for faults 5, 10, 11, 16, 20, and 21 is not ideal. Fault 5 is a step change in the feed temperature of the compressor condensate. In the initial stage of the fault, the variables affect each other, and the occurrence of the fault can be detected through the mutual information matrix mapping. However, since the control system then responds and makes corresponding compensations, the system presents problems under the new operating conditions, and the features extracted based on the mutual information projection are below the control limit, reflecting the dynamic regulation ability of the system. Additionally, for faults 10, 11, 20, and 21, the dynamic monitoring indicators of the second layer To a certain extent, the occurrence of faults is detected, and the time-series changes in the mutual information between variables reflect that these four faults are disturbances with continuous dynamic anomalies, which is an effective supplement to the fault detection performance.

[0130] Table 2 summarizes the fault detection rates of different methods for 17 fault types in the TE process. Among them, the bolded group is the optimal group for that fault. The first and second rows under each fault correspond to the false alarm rate and the fault detection rate respectively.

[0131] Table 2:

[0132]

[0133] The embodiments of the present invention have a good recognition effect on the process dynamic performance monitoring of the process. Taking Fault 14 as an example, the dynamic detection ability of the present invention for faults is analyzed.

[0134] Fault 14 refers to the stickiness of the reactor cooling water valve, and the fault occurs at the 161st sample. Fault 14 is a more complex fault case, which refers to the stickiness of the reactor cooling water valve. When Fault 14 occurs, the cooling water valve fluctuates severely. This will directly cause the reactor cooling water flow rate, reactor temperature, reactor cooling water outlet temperature, and reactor pressure to oscillate simultaneously, and indirectly cause small fluctuations in the reactor pressure, separator pressure, and stripper pressure. Fault 14 cannot be eliminated by the control system, and these variables will remain in an abnormal state continuously.

[0135] For Fault 14, Figure 2 、 Figure 3 、 Figure 4 respectively present the monitoring results based on the PCA method, the SFA method, and the embodiments of the present invention. As can be seen from Figure 2 Visible, the fault detection indexes T 2 and SPE based on PCA can successfully detect the fault, but the fault false alarm rate is relatively high; as can be seen from Figure 3 Visible, only statistic in the SFA-based method is effective, the false alarm rate of the statistic is too high, but the statistic fluctuates above and below the threshold and cannot clearly indicate that this fault is a fault that cannot be eliminated by the control system; as can be seen from Figure 4 Visible, the monitoring indexes T 2 and mainly for fault detection in the embodiments of the present invention can both successfully identify the changes in operating conditions, and even the statistic also shows good detection performance; further, After the fault 14 was detected by the statistics, it has been in a high-alarm state, indicating that the control loop cannot eliminate the fault through regulation, which is consistent with the actual situation and verifies the comprehensive dynamic monitoring ability of the embodiment of the present invention for the control system.

[0136] Embodiment 2

[0137] Corresponding to the method embodiment, in the embodiment of the present invention, a hierarchical multivariate process monitoring device based on mutual information matrix projection is further provided, such as a terminal, a server, etc. Among them, the server can be an independent physical server, or a server cluster or distributed system composed of multiple physical servers, or a cloud server providing basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communications, middleware services, domain name services, security services, CDN, and big data and artificial intelligence platforms. The terminal can be a smart phone, a tablet computer, a notebook computer, a desktop computer, etc., but is not limited thereto.

[0138] An example diagram of the hardware structure block diagram of the hierarchical multivariate process monitoring device based on mutual information matrix projection provided by the embodiment of the present invention is as Figure 5 shown, and may include:

[0139] Processor 1, communication interface 2, memory 3, and communication bus 4;

[0140] Among them, the processor 1, the communication interface 2, and the memory 3 complete mutual communication through the communication bus 4;

[0141] Optionally, the communication interface 2 can be an interface of a communication module, such as an interface of a GSM module;

[0142] The processor 1 may be a central processing unit CPU, or a specific integrated circuit ASIC (Application Specific Integrated Circuit), or one or more integrated circuits configured to implement the embodiments of the present application.

[0143] The memory 3 may include high-speed RAM memory, and may also include non-volatile memory, such as at least one disk memory.

[0144] Among them, the processor 1 is specifically configured to execute the computer program stored in the memory 3 to perform the following steps:

[0145] Its offline modeling steps include:

[0146] S11. Perform standardized preprocessing on the original data of the time series data in discrete form to generate a standard training set for modeling;

[0147] S12. Perform the first - layer feature projection based on the mutual - information matrix, including: approximately calculating the mutual information between variables for the standard training set using the matrix Rényi's α - entropy function, constructing a mutual - information matrix, and obtaining the corresponding first - eigenvector matrix and first - eigenvalue matrix through eigenvalue decomposition of the mutual - information matrix; extracting the first - transformation - element feature through spatial projection based on the mutual - information matrix;

[0148] S13. Determine the principal - component subspace and the residual subspace, including: determining the number of elements in the principal - component space according to a preset eigenvalue contribution rule, denoted as the number of principal components; and obtaining the corresponding principal - eigenvector matrix, residual - eigenvector matrix, principal - eigenvalue matrix, and residual - eigenvalue matrix;

[0149] S14. Determine the first control limit for fault detection of the first - layer feature projection according to the monitoring statistic and confidence level of the standard training set; the first control limit is used as a fault - alarm threshold;

[0150] S15. Perform the second - layer feature projection based on the mutual - information matrix, including: constructing a differential space with the first - transformation - element feature at a time delay τ; calculating the time - lagged mutual - information matrix on the differential space; performing eigenvalue decomposition on the time - lagged mutual - information matrix to obtain a second - eigenvector matrix and a second - eigenvalue matrix, and obtaining the first slow - mutual - information feature based on the time - lagged mutual - information matrix;

[0151] S16. Determine the principal - component subspace and the residual subspace of the slow - mutual - information feature projection, including: obtaining the principal - eigenvector matrix, residual - eigenvector matrix, principal - eigenvalue matrix, and residual - eigenvalue matrix of the slow - mutual - information feature projection according to the contribution rule of the slowness self - mutual information;

[0152] S17. Determine multiple second control limits for dynamic monitoring of the second - layer feature projection, including: calculating each of the second control limits according to the covariance matrix of the principal - slow - mutual - information feature and the residual - slow - mutual - information feature, and the covariance matrix of the first - order difference of the principal - slow - mutual - information feature and the residual - slow - mutual - information feature; each of the second control limits is used to judge the dynamic characteristics of the system control performance after a fault.

[0153] Preferably, in the embodiment of the present invention, it further includes a step of online monitoring, including:

[0154] S21. Standardize and pre - process the online monitoring data using the mean and variance of the original training set to generate a standard test set for model input;

[0155] S22. Perform the extraction of meta-features of the test set conversion based on the mutual information matrix, including: projecting the standard test set onto the space based on the mutual information matrix to obtain the second meta-features of the test set conversion based on the mutual information matrix;

[0156] S23. Perform the extraction of slow mutual information features of the differential space based on the mutual information matrix, including: performing differential calculation on the second meta-features with a time delay τ, and projecting the differential features onto the time-delay mutual information matrix space of the standard training set to obtain the second slow mutual information features;

[0157] S24. Calculate the online monitoring indicators, including: calculating the first monitoring indicator of the projection of the first-layer features of the standard test set, and a plurality of second monitoring indicators on the principal component subspace and the residual subspace of the projection of the second-layer features of the standard test set;

[0158] S25. Use the first control limit as a threshold to judge the fault alarm according to the first monitoring indicator, and use each of the second monitoring indicators as parameters to monitor the system control performance according to each of the second control limits, and judge the dynamic characteristics of the system after a fault.

[0159] The above product can execute the method provided by the embodiment of the present invention, and has the corresponding functional modules and beneficial effects of the executed method. For the technical details not described in detail in this embodiment, reference can be made to the hierarchical multivariate process monitoring method based on the mutual information matrix projection provided by the embodiment of the present invention.

[0160] The above product can execute the method provided by the embodiment of the present invention, and has the corresponding functional modules and beneficial effects of the executed method. For the technical details not described in detail in this embodiment, reference can be made to the methods provided by other embodiments of the present invention.

[0161] Those of ordinary skill in the art can realize that the units and algorithm steps of each example described in combination with the embodiments disclosed herein can be implemented by electronic hardware, or by a combination of computer software and electronic hardware. Whether these functions are executed in hardware or software depends on the specific application and design constraints of the technical solution. Professional technicians can use different methods to implement the described functions for each specific application, but such implementation should not be considered to exceed the scope of this application.

[0162] In several embodiments provided in the present application, it should be understood that the disclosed systems, devices, and methods can be implemented in other ways. On the other hand, the couplings or direct couplings or communication connections shown or discussed with each other can be through some interfaces, and the indirect couplings or communication connections of devices or units can be in electrical, mechanical, or other forms.

[0163] The unit described as a separation component may or may not be physically separated. The component shown as a unit may or may not be a physical unit, that is, it may be located in one place or may be distributed to multiple network units. Some or all of the units can be selected according to actual needs to achieve the purpose of the solution of this embodiment.

[0164] In addition, in each embodiment of the present application, each functional unit may be integrated in a processing unit, may exist physically alone for each unit, or two or more units may be integrated in one unit.

[0165] It should be understood that in the embodiments of the present application, dependent claims, each embodiment, and features can be combined with each other to solve the foregoing technical problems.

[0166] If the said function is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on such an understanding, the technical solution of the present application, in essence, or the part that contributes to the prior art, or a part of this technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions for causing a computer device (which may be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the methods described in each embodiment of the present application. The foregoing storage medium includes: various media such as USB flash drives, mobile hard disks, read-only memories (ROM, Read-Only Memory), random access memories (RAM, Random Access Memory), magnetic disks, or optical discs that can store program codes.

[0167] The above description of the disclosed embodiments enables those skilled in the art to implement or use the present application. Various modifications to these embodiments will be obvious to those skilled in the art. The general principles defined herein can be implemented in other embodiments without departing from the spirit or scope of the present application. Therefore, the present application will not be limited to these embodiments shown herein, but will be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A hierarchical multivariate process monitoring method based on mutual information matrix projection, characterized in that The steps of its offline modeling include: S11. Perform standardized preprocessing on the original data of discrete time-series data to generate a standard training set for modeling; S12. Perform the first-layer feature projection based on the mutual information matrix, including: approximately calculating the mutual information between variables for the standard training set using the α-entropy function based on matrix Rényi, constructing a mutual information matrix, and obtaining the corresponding first eigenvector matrix and first eigenvalue matrix through eigen-decomposition of the mutual information matrix; extracting the first transformed meta-feature through spatial projection based on the mutual information matrix; S13. Determine the principal component subspace and the residual subspace, including: determining the number of elements in the principal component space according to the preset eigenvalue contribution degree rule, denoted as the number of principal components; and obtaining the corresponding principal component eigenvector matrix, residual eigenvector matrix, principal component eigenvalue matrix, and residual eigenvalue matrix; S14. Determine the first control limit for fault detection of the first-layer feature projection according to the monitoring statistic and confidence level of the standard training set; the first control limit is used as the fault alarm threshold; S15. Perform the second-layer feature projection based on the mutual information matrix, including: constructing a differential space using the first transformed meta-feature with a time delay τ; calculating the time-delay mutual information matrix on the differential space; performing eigen-decomposition on the time-delay mutual information matrix to obtain the second eigenvector matrix and the second eigenvalue matrix, and obtaining the first slow mutual information feature based on the time-delay mutual information matrix; S16. Determine the principal component subspace and the residual subspace of the slow mutual information feature projection, including: obtaining the principal component eigenvector matrix, residual eigenvector matrix, principal component eigenvalue matrix, and residual eigenvalue matrix of the slow mutual information feature projection according to the contribution degree rule of the slowness self-mutual information; S17. Determine multiple second control limits for dynamic monitoring of the second-layer feature projection, including: calculating each of the second control limits according to the covariance matrix of the principal component slow mutual information feature and the residual slow mutual information feature, and the covariance matrix of the first-order difference of the principal component slow mutual information feature and the residual slow mutual information feature; each of the second control limits is used to judge the dynamic characteristics of the system control performance after a fault.

2. The hierarchical multivariate process monitoring method based on mutual information matrix projection according to claim 1, characterized in that It also includes the steps of online monitoring, including: S21. Perform standardized preprocessing on the online monitoring data using the mean and variance of the original training set to generate a standard test set for model input; S22. Extract the test set transformed meta-feature based on the mutual information matrix, including: projecting the standard test set onto the space based on the mutual information matrix to obtain the second transformed meta-feature based on the mutual information matrix; S23. Extract the differential space slow mutual information feature based on the mutual information matrix, including: performing a differential calculation on the second transformed meta-feature with a time delay τ and projecting the differential feature onto the space of the time-delay mutual information matrix of the standard training set to obtain the second slow mutual information feature; S24. Calculate the online monitoring indicators, including: calculating the first monitoring indicator of the first-layer feature projection of the standard test set, and multiple second monitoring indicators on the principal component subspace and the residual subspace of the second-layer feature projection of the standard test set; S25. Judge fault alarms based on the first monitoring index with the first control limit as the threshold, and monitor the system control performance based on each second monitoring index according to each second control limit to judge the dynamic characteristics of the system after a fault.

3. The hierarchical multivariate process monitoring method based on mutual information matrix projection according to claim 2, characterized in that The preprocessing of the original data of the discrete-form time series data into a standardized training set for modeling includes: Construct an original training set for the original process data A standardized training set is obtained by standardizing the original training set according to the following formula Among them, is the mean of the original training set, is the standard deviation of the original training set.

4. The hierarchical multivariate process monitoring method based on mutual information matrix projection according to claim 3, wherein Using the standardized training set to calculate the mutual information between variables by means of the α-entropy function based on matrix Rényi, constructing a mutual information matrix, and performing eigenvalue decomposition on the mutual information matrix to obtain the corresponding first eigenvector matrix and first eigenvalue matrix; Performing spatial projection extraction based on the mutual information matrix to obtain the first transformation element features, including: The mutual information matrix based on the α-entropy function of matrix Rényi, including: Define the kernel function κ: Calculate the sensor variable x in the standardized test set i and x j of the Gram matrix K = κ(x i , x j ); Normalize the Gram matrix to obtain Calculating the α-entropy function based on matrix Rényi according to the matrix A; Among them, represents the Hadamard product between matrices A and B; Construct a variable mutual information matrix based on the matrix Rényi α-entropy function, where the mutual information matrix Specifically includes: Among them, H(x i ) is the entropy of variable x i , H(x i ; x j ) is the joint entropy between variable x i and x j , I(x i ; x j ) is the mutual information between variable x i and x j , and I(x i ; x j ) = H(x i ) + H(x j ) - H(x i ; x j ); In particular, on the diagonal of the mutual information matrix, there is I(x i ; x i ) = H(x i ); Perform eigen - decomposition on the mutual information matrix to obtain the first eigen - vector matrix and the first eigenvalue matrix Λ = diag(λ1, λ2…, λ ) ∈ R m m×m ;​ The first transformation element feature Z is extracted by performing spatial projection based on the mutual information matrix: train ={z train (1), z train (2), …, z train (n)}, 5. The hierarchical multivariate process monitoring method based on mutual information matrix projection according to claim 4, wherein Determining the number of elements in the principal component space according to the preset eigenvalue contribution rule, denoted as the number of principal components; And obtaining the corresponding principal component eigenvector matrix, residual eigenvector matrix, principal component eigenvalue matrix and residual eigenvalue matrix, including: According to the eigenvalue contribution rule Determine the number of elements in the principal component space, denoted as the principal component number I, and obtain the principal component eigenvector matrix W I , the residual eigenvector matrix W m-I , the principal component eigenvalue matrix Λ I and the residual eigenvalue matrix Λ m-I .

6. The hierarchical multivariate process monitoring method based on mutual information matrix projection according to claim 5, wherein Determining the first control limit for fault detection of the first-layer feature projection according to the monitoring statistic and confidence level of the standardized training set, including: Determine the fault detection control limit for the first layer feature projection Denote it as the first control limit: Calculate the statistical index T of the standard training set data 2 , and obtain it at the confidence level a T 2 = x T UΛ -1 U T x; where U represents the mutual information matrix of the eigenvector matrix, and Λ represents the corresponding eigenvalue matrix, obtained at a confidence level of a 7. The hierarchical multivariate process monitoring method based on mutual information matrix projection according to claim 6, characterized in that Using the first transformation element features to construct a differential space with a time delay τ; Calculating the time-delay mutual information matrix on the differential space; Performing eigenvalue decomposition on the time-delay mutual information matrix to obtain a second eigenvector matrix and a second eigenvalue matrix, and obtaining the first slow mutual information feature based on the time-delay mutual information matrix, including: Projection of the first transformed eigenfeature space projection Z after projection train = {z train (1), z train (2), …, z train (n)}, taking the difference space with time delay τ Calculate the time-delay mutual information matrix on the differential space Among them, is the mutual information between and ; Decompose into feature components to obtain a second feature vector matrix and a second eigenvalue matrix Obtaining the first slow mutual information feature based on the time-delay mutual information matrix; Y train = {y train (1), y train (2), …, y train (n)}.

8. The hierarchical multivariate process monitoring method based on mutual information matrix projection according to claim 7, characterized in that Obtaining the principal component eigenvector matrix, residual eigenvector matrix, principal component eigenvalue matrix and residual eigenvalue matrix of the slow mutual information feature projection according to the contribution rule of the slowness self-mutual information, including: Calculate the self - mutual information AI(y 1:n-τ ; y 1+τ:n ) of the first slow mutual - information feature. According to the contribution - degree rule of the slowness self - mutual information screen the principal components, denoted as the number of principal components d, and obtain the principal - component feature matrix W T,d of the slow mutual - information feature projection, the residual - feature matrix W T,e , the principal - component eigenvalue matrix Ω d and the residual - eigenvalue matrix Ω e .

9. The hierarchical multivariate process monitoring method based on mutual information matrix projection according to claim 8, wherein, Calculating each second control limit according to the covariance matrix of the principal component slow feature and the residual slow feature, and the covariance matrix of the first-order difference of the principal component slow feature and the residual slow feature, including: The control limit for the dynamic monitoring of the second-layer feature projection is denoted as the second control limit, including and Use and to calculate and statistics; use and to calculate and statistics; where, Σ d and Σ e are the covariance matrices of the slow mutual information features y d of the principal subspace and the slow mutual information features y e of the residual subspace respectively; Π d and Π e are the covariance matrices of the first-order differences of the principal slow mutual information features and the first-order differences of the residual slow mutual information features respectively; The second control limit is calculated at a confidence level of a and 10. The hierarchical multivariate process monitoring method based on mutual information matrix projection according to claim 9, wherein Preprocessing the online monitoring data into a standardized test set for model input, including: Constructing an original data test set according to the online monitoring data in the chemical process; Using the mean of the original training set and variance preprocess the test set by standardization to obtain a standard test set 11. The hierarchical multivariate process monitoring method based on mutual information matrix projection according to claim 10, characterized in that The second conversion element feature includes: z(t) = Λ -12 Ux(t); The second slow mutual information feature includes: y(t) = P T z(t) = P T Λ -12 Ux(t); The first monitoring index includes: T 2 = x T UΛ -1 U T x; The second monitoring indicator includes: and Wherein: In the above formulas, U represents the mutual information matrix of the training set of the eigenvector matrix, and Λ represents the corresponding eigenvalue matrix; y d and y e respectively represent the slow mutual information features of the principal component subspace and the residual subspace of the test set, Σ d and Σ e respectively represent the covariance matrices of y d and y e , and are respectively the first-order differences of the slow mutual information features of the principal components and the first-order differences of the slow mutual information features of the residuals of the covariance matrix.

12. A hierarchical multivariate process monitoring device based on mutual information matrix projection, characterized in that, including: A memory for storing a computer program; A processor for calling and executing the computer program to implement the steps of the hierarchical multivariate process monitoring method based on mutual information matrix projection according to any one of claims 1-11.

Images