A knowledge- and data-driven distributed process monitoring method for plants

Abstract

The present invention discloses a plant-level distributed process monitoring method based on dual drive of knowledge and data. First, the causal relationship between variables is determined through process knowledge in the process flow, and a knowledge causal graph is obtained through the causal relationship between variables. The process flow is modeled through the knowledge causal graph, and the variables in the process are divided into several physically meaningful sub-blocks, so that the process variable decomposition is interpretable. However, if the causal relationship between variables is used as the only indicator to measure the block, it cannot guarantee that the subsequent fault detection performance reaches the optimal level. To this end, based on the knowledge causal graph block, standardized mutual information is introduced to obtain the data correlation between variables in the sub-block, and the variables with high correlation are further divided into separate sub-blocks. The final division result is both interpretable and more correlated in data. The method of first local detection and then global fusion is used in the fault detection stage, which has the advantage of global problem discovery and rapid location of sub-blocks.

Description

Technical Field

[0001] The present invention belongs to the field of industrial process monitoring and fault detection, and in particular relates to a plant-level distributed process monitoring method based on dual drive of knowledge and data. Background Art

[0002] With the development of plant-level processes, factories are becoming larger and larger, and production processes are becoming increasingly complex. To address this phenomenon, distributed process monitoring technology has been developed. Distributed monitoring primarily involves process variable decomposition, local monitoring design, and the integration of decision-making systems. Process variable decomposition is crucial in distributed monitoring, and the rationality of the decomposition impacts subsequent detection performance. There are two main types of process variable decomposition: knowledge-based and data-driven. Knowledge-based process decomposition emphasizes strong causal relationships between variables within sub-blocks rather than detection precision, improving interpretability but limiting detection accuracy. In contrast, data-driven process decomposition relies on variable data correlation to achieve high detection rates, but neglects the interpretability of process monitoring. Summary of the Invention

[0003] In view of the shortcomings of the existing technology, the purpose of the present invention is to propose a plant-level distributed process monitoring method based on both knowledge and data dual drive, so that the variable decomposition is both interpretable and more relevant, ensuring the performance of subsequent fault detection and diagnosis while also improving the interpretability.

[0004] The present invention is achieved through the following technical solutions:

[0005] A knowledge- and data-driven distributed process monitoring method for a plant includes the following steps:

[0006] (1) Collect the data set of normal operating conditions of the process, that is, the training set, denoted as X m×n , m is the number of variables, n is the number of samples;

[0007] (2) According to the attributes of the variables in the process flow diagram, the knowledge causal graph G = (C, E) is determined, where C is the node set of the dependent variable and E is the node set of the result variable. The data set is divided into several physically meaningful sub-blocks by knowledge causality and recorded as X = [X1, X2, ..., X q ];

[0008] (3) Calculate the normalized mutual information value NMI between variables in the sub-block ij , and set a threshold NMI il , NMI ij >NMI il The variables are further divided into separate sub-blocks denoted as X b Represents the b-th sub-block after division;

[0009] (4) Using random kernel principal component analysis to determine the threshold of each sub-block is recorded as and

[0010] (5) For the online detection stage, the online collected data Y m×t Steps (2) to (3) are divided into blocks and recorded as

[0011] (6) Perform random kernel principal component analysis on each sub-block to obtain the detection statistic and

[0012] (7) If or A fault is detected if and There is no fault;

[0013] (8) The detection results of each sub-block are Bayesian fused to obtain the global fault detection statistics and SPE GFI , the significance level α is recorded as the global detection control limit; if or SPE GFI >α will be displayed as a fault. And SPE GFI ≤α is normal state;

[0014] (9) For the detected fault, the contribution rate between sub-blocks is calculated first and recorded as CBCI b ,The sub-block with high contribution rate is defined as the faulty sub-block and recorded as FBS;

[0015] (10) Perform variable contribution analysis on the fault sub-block, denoted as CVCI b , the variable with high contribution rate is defined as the fault variable and recorded as FVS;

[0016] (11) Use the knowledge causal graph to perform causal analysis on FBS and FVS to obtain the root cause and propagation path of the fault.

[0017] The step (2) is: knowledge causal graph G = (C, E), where C = {c1, c2, ..., c g} and E={e1,e2,…,e f} distribution represents the node set of dependent variables and outcome variables, G∈R g×f It is a knowledge causal matrix, where both rows and columns are represented as nodes; the elements in the matrix are composed of 0 and 1, where 1 represents a causal relationship between two variables and 0 represents no causal relationship.

[0018] The step (3) is as follows: mutual information is to measure the correlation between variables by using the entropy relationship between two variables. The greater the correlation between the variables, the greater the value of mutual information. Mutual information is measured by the difference between entropy and conditional entropy, which is expressed as:

[0019] I(X;Y)=H(X)-H(X|Y)

[0020]

[0021]

[0022] Normalized mutual information means limiting the value of mutual information to between 0 and 1, which is expressed as: The threshold is set empirically as follows:

[0023] The step (4) is as follows: random kernel principal component analysis is to project the high-dimensional kernel matrix into a low-rank approximate random subspace, to ensure that the eigenvalues ​​and eigenvectors are approximately unchanged to the greatest extent; for the b-th sub-block Use radial basis kernel function to calculate kernel matrix K b : The centralization of the kernel matrix is ​​recorded as: I n×n ∈R n×n The elements are Before calculating the threshold, you need to implement the random projection algorithm: set a random matrix Ω b ∈R n×l , l≤n, calculate Y b Perform QR decomposition to obtain Q b , Q b satisfy ε is the approximation error, The matrix Limited to B b In this subspace, for B b Perform feature decomposition Among them L b The diagonal is the eigenvalue Matrix sorted from large to small; by Q b U b get Is an approximate eigenvector matrix containing l principal eigenvalues, for the eigenvector Normalize to: Determine the nonlinear principal component of the training data as For the principal space p b The determination of is determined by the cumulative variance percentage; the detection threshold with a significance level α is expressed as:

[0024]

[0025] in, It is p b and Np b F distribution composed of degrees of freedom; It is h b Chi-square distribution with degrees of freedom, g b Is a weight parameter; assuming that the training sample The mean and variance of b and b b ,but

[0026] The step (6) is: for the b-th sub-block The kernel matrix K 1b The radial basis kernel function is used as The centralization of the kernel matrix is ​​recorded as: The elements are Determine the nonlinear principal component of online data as: The test statistic is recorded as:

[0027]

[0028] in, Λ b Front p b A diagonal matrix consisting of eigenvalues.

[0029] The step (8) is as follows: the Bayesian fusion strategy includes two steps: probability conversion and probability fusion; for the b-th sub-block, firstly, the sub-block monitoring statistics are converted into the sub-block failure probability:

[0030] in The calculation is the same as that of the prior probability, which represents the marginal probability, N and F represent normal and abnormal situations; The likelihood probability is expressed as: and and The calculation is the same; then the sub-block failure probabilities are fused to obtain the global monitoring statistics, which is specifically recorded as:

[0031]

[0032] The significance level α is recorded as the global detection control limit; if or SPE GFI >α will be displayed as a fault. And SPEGFI ≤α is the normal state.

[0033] The step (9) is: analyzing the fault sub-blocks of the detected fault, assuming that n F fault samples, the principal component space contribution index TBCI of the b-th sub-block b and residual space contribution index QBCI b Denoted as: when when Define the combined sub-block contribution index CBCI b Denoted as: The sub-blocks with larger contribution rates are located as the faulty block set.

[0034] The step (10) is: calculating the variable contribution index of the fault block, and the calculation steps for the b-th fault sub-block are:

[0035]

[0036]

[0037]

[0038]

[0039]

[0040] where j = 1, ..., m b , m b is the number of variables in sub-block b, Θ is a fault experience parameter, I 1n =[1,···,1]∈R 1×n , I nn ∈R n×n ;Will with CVSPE b Combine to obtain CVCI b : The variables with larger contribution rates are defined as the fault variable set.

[0041] The step (11) is: combining the FBS and FVS obtained in steps (9) and (10) with the knowledge causal graph, performing causal analysis on them, and identifying the root cause and propagation path of the fault; first, at the sub-block level, identifying the causal relationship between the fault sub-blocks based on the knowledge causal graph; the variable set of sub-block B1 is V1, the variable set of sub-block B2 is V2, and the knowledge causal graph between V1 and V2 is represented by G 12 , G 12 The internal node set is V c, the result node set is represented as V e ; The causal relationship between B1 and B2 is:

[0042]

[0043] Where B1→B2 means B1 is the fault cause block. Represents the fault sub-blocks of B1 and B2 that are causally related to each other; the causal relationship between the fault sub-blocks is obtained: G FBS =(C FBS ,E FBS ), select the sub-block without node pointing as the root sub-block; secondly, at the variable level, the causal relationship between fault variables is G FVS =(C FVS ,E FVS ), through G FVS The source node is used to identify the root cause of the fault.

[0044] The beneficial effects of the present invention are:

[0045] The present invention realizes the variable decomposition in the process variable decomposition stage in the distributed monitoring framework, which is both explainable and has higher variable correlation, thereby improving the performance and interpretability of subsequent fault detection and diagnosis. Specifically, for the plant-level process, a knowledge causal graph is first proposed to model the process, and the process variables are divided into several sub-blocks with practical significance through the knowledge causal graph, so that the sub-blocks have interpretability. On this basis, it is proposed to analyze the variable correlation of the sub-blocks with physical significance, and thus the standardized mutual information is introduced to further analyze the variables in the sub-blocks, and the variables with higher correlation are further divided. Through the above steps, the variable division results are both explainable and have higher correlation; the divided sub-blocks are subjected to local random kernel principal component analysis, and the local detection results are globally fused using the Bayesian fusion strategy, forming an effect that can detect both locally and globally; for the detected faults, a method combining knowledge and data is used to perform fault tracing analysis, which improves the interpretability of the diagnosis results and reduces the smearing effect. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] Figure 1 It is a flow chart of the present invention.

[0047] Figure 2 It is the modeling of the Tennessee-Eastman process by the knowledge causal diagram.

[0048] Figure 3 It is a variable decomposition of the Tennessee-Eastman process based on the knowledge causal diagram.

[0049] Figure 4 It is a variable decomposition of the Tennessee-Eastman process using standardized mutual information based on knowledge partitioning.

[0050] Figure 5 It is a knowledge- and data-driven, plant-level distributed process monitoring method for the local and global detection results of Tennessee-Eastman process fault 6.

[0051] Figure 6 It is a knowledge- and data-driven, dual-driven, plant-level distributed process monitoring method that measures the contribution of the six sub-blocks and variables of the Tennessee-Eastman process fault.

[0052] Figure 7 It is a knowledge- and data-driven, plant-level distributed process monitoring method that analyzes the propagation paths of the six sub-blocks and variables of the Tennessee-Eastman process fault. DETAILED DESCRIPTION

[0053] The present invention is further described below with reference to the accompanying drawings and embodiments.

[0054] The present invention provides a plant-level distributed process monitoring method based on knowledge and data dual drive, such as Figure 1 As shown, the steps are as follows:

[0055] The first step is to collect a data set of normal operating conditions of the process, that is, the training set, denoted as X m×n , m is the number of variables, and n is the number of samples.

[0056] The second step is to determine the knowledge causal graph G=(C,E) according to the attributes of the variables in the process flow diagram, where C={c1,c2,…,c g} is the node set of dependent variables, E={e1,e2,…,e f} is the node set of the result variable, G∈R g×f The knowledge causal matrix is ​​represented by rows and columns as nodes. The elements in the matrix are composed of 0 and 1, 1 represents the causal relationship between the two variables, and 0 represents no causal relationship. The data set is divided into several physically meaningful sub-blocks through the knowledge causal graph, denoted as X = [X1, X2, ..., X q ].

[0057] The third step is mutual information, which uses the entropy relationship between two variables to measure the correlation between them. The greater the correlation between the variables, the greater the mutual information value. Mutual information is measured by the difference between entropy and conditional entropy, which is expressed as:

[0058] I(X;Y)=H(X)-H(X|Y)

[0059]

[0060]

[0061] The normalized mutual information refers to restricting the value of the mutual information between 0 and 1, expressed as: The threshold is determined by the empirical method and set as: Calculate the normalized mutual information value between variables within the sub-block and further divide the variables with ij >NMI il into separate sub-blocks denoted as X b which represents the b-th sub-block divided by the knowledge causal graph.

[0062] In the fourth step, the random kernel principal component analysis projects the high-dimensional kernel matrix into a low-rank approximate random subspace, ensuring that the eigenvalues and eigenvectors are approximately invariant to the greatest extent. For the b-th sub-block the radial basis kernel function is used to calculate the kernel matrix K b : The centering process of the kernel matrix is denoted as: I n×n ∈R n×n and the elements in it are Before calculating the threshold, the implementation of the random projection algorithm is required: Set a random matrix Ω b ∈R<​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​ It is h b Chi-square distribution with degrees of freedom, g b Is a weight parameter; assuming that the training sample The mean and variance of b and b b ,but

[0065] The fifth step is the online detection phase, which involves collecting data Y online. m×t The steps from the second step to the third step are divided into blocks and recorded as

[0066] Step 6: For the b-th sub-block The kernel matrix K 1b The radial basis kernel function is used as The centralization of the kernel matrix is ​​recorded as: I t×n ∈R t×n The elements are Determine the nonlinear principal component of online data as: The test statistic is recorded as:

[0067]

[0068] in, For L b Front p b A diagonal matrix consisting of eigenvalues.

[0069] Step 7: If or A fault is detected if and No fault.

[0070] In the eighth step, the detection results of each sub-block are globally integrated. The Bayesian fusion strategy includes two steps: probability conversion and probability fusion. For the b-th sub-block, the sub-block monitoring statistics are first converted into sub-block failure probability:

[0071] in The calculation of is the same, both represent marginal probability, N and F represent normal and abnormal situations. Prior probability The likelihood probability is expressed as: and and and The calculation is the same. Then the sub-block failure probabilities are fused to obtain the global monitoring statistics, which is specifically recorded as:

[0072]

[0073] The significance level α is recorded as the global detection control limit. or SPE GFI >α will be displayed as a fault. And SPE GFI ≤α is the normal state.

[0074] The ninth step is to analyze the fault sub-blocks of the detected faults. Assuming that n F fault samples, the principal component space contribution index TBCI of the b-th sub-block b and residual space contribution index QBCI b Denoted as:

[0075]

[0076] Because the principal component space contribution index Contains Therefore, the contributions of the principal component space are sensitive to each other. In order to prevent misjudgment, the combined sub-block contribution index CBCI is defined. b Denoted as:

[0077]

[0078] The sub-blocks with larger contribution rates are located as the fault block set (Fault Block Set, FBS).

[0079] The tenth step is to calculate the variable contribution index of the faulty sub-block. The calculation steps for the bth faulty sub-block are:

[0080]

[0081]

[0082]

[0083]

[0084]

[0085] where j = 1,…,m b , m b is the number of variables in sub-block b, Θ is a fault experience parameter, I 1n =[1,…,1]∈R 1×n , I nn ∈R n×n .Will with CVSPE b Combine to obtain CVCIb : The variables with larger contribution rates are defined as the fault variable set (Fault Variable Set, FVS).

[0086] In the eleventh step, the FBS and FVS obtained in the ninth and tenth steps are combined with the knowledge causal graph to perform causal analysis and identify the root cause and propagation path of the fault. First, at the sub-block level, the causal relationship between the fault sub-blocks is identified based on the knowledge causal graph. For example: the variable set of sub-block B1 is V1, the variable set of sub-block B2 is V2, and the knowledge causal graph between V1 and V2 is represented by G 12 , G 12 The internal node set is V c , the result node set is represented as V e The causal relationship between B1 and B2 is:

[0087]

[0088] Where B1→B2 means B1 is the fault cause block. Represents the fault sub-blocks of B1 and B2 that are causally related to each other. The causal relationship between the fault sub-blocks is obtained: G FBS =(C FBS ,E FBS ), select the sub-block without node pointing as the root sub-block. Secondly, at the variable level, the causal relationship between fault variables is G FVS =(C FVS ,E FVS ), through G FVS The source node is used to identify the root cause of the fault.

[0089] Application Examples

[0090] The effectiveness of the proposed method is demonstrated below using a specific Tennessee Eastman (TE) process. The process variables are 52 (22 continuously measured variables, 19 component measured variables, and 11 controlled variables). The continuously measured variables are sampled every three minutes, while the component variables require a certain delay. To better match the actual process, the proposed method uses a total of 33 continuously measured and controlled variables for experimental demonstration. 500 samples were collected under normal operating conditions, and 960 fault samples were collected online (the first 160 of which are normal samples). Fifteen fault sets were selected for a detailed description of Fault 4.

[0091] Next, the implementation steps of the present invention are described in detail in conjunction with this specific process:

[0092] 1. Standardize the data collected under normal working conditions and reflect the causal relationship between process variables based on process knowledge. Table 1 is the knowledge causal matrix of TE process variables. Convert it into a knowledge causal graph to model the process flow chart, such as Figure 2 For example, v1 and v 44 are mutually causal variables, v6 is the dependent variable of v7, and the process variables are divided into several sub-blocks with physical meanings, such as Figure 3 For example, sub-block 1 is shown as the feed zone variables and some process variables inside the reactor.

[0093] 2. For the divided sub-blocks, the normalized mutual information is used to calculate the correlation between the variables in the sub-blocks. The variables whose correlation exceeds the threshold are further divided into a sub-block. The final division result is both interpretable and more relevant, such as Figure 4 As shown, for example, sub-block 1 represents the process of feed A, and sub-block 7 includes the circulation reaction in the stripper and part of the internal reaction.

[0094] 3. Use random kernel principal component analysis to perform principal component space analysis on all sub-blocks and residual space Calculation.

[0095] 4. Perform steps 1 and 2 on the fault sample data collected online and divide it into blocks. Use random kernel principal component analysis to perform monitoring statistics for each sub-block. and If or A fault is detected if and There is no fault. Table 2 shows the monitoring statistics of the sub-block. The first row shows the average false alarm rate of 15 faults. Each bold row indicates the highest detection result under this fault. The attribute of fault 6 is the feed loss of material A. When the operating valve v 44 If a failure occurs, it will result in the loss of material A, then the root cause of failure 6 is v 44 , in sub-block 1, its local detection results are as follows Figure 5 As shown in part (b), T 2 Both the SPE and the MCU can detect faults.

[0096] 5. Use the Bayesian fusion method to globally integrate the local monitoring results to form global monitoring statistics and SPE GFI , the significance level α is recorded as the global detection control limit. or SPE GFI >α will be displayed as a fault. And SPE GFI≤α is a normal state. Table 3 shows the comparison of the detection results between the global fusion detection results DPM-KCDMI and KCD-RKPCA and MI-RKPCA, where KCD-RKPCA is the global fusion detection result of the process based only on the knowledge causal graph block, and MI-RKPCA is the global fusion detection result of the process based only on the mutual information block. The numbers in the table represent the difference between the detection rate and the false alarm rate. Each bold line represents the highest performance under this fault. The global detection results of fault 6 are as follows: Figure 5 As shown in part (a), T 2 Both the SPE and the MCU can detect the fault.

[0097] 6. For the detected fault, first calculate the contribution rate between sub-blocks and record it as CBCI b , define the sub-block with high contribution rate as the fault sub-block and record it as FBS. The contribution rate of the sub-block for fault 6 is as follows Figure 6 As shown in part (a), sub-blocks 1, 2, 3, 5, and 7 are recorded as FBS. Then, the variable contribution analysis within the fault sub-block is performed and recorded as CVCI. b , define the variable with high contribution rate as fault variable and record it as FVS. Figure 6 (b-1), (b-2), (b-3), (b-4), (b-5) are the variable contribution indicators in the corresponding fault sub-block. 44 ,v 51 ,v7,v 13 ,v 46 ,v 16} is selected as the fault variable. Finally, the knowledge causal graph is used to perform causal analysis on FBS and FVS to obtain the fault root cause and propagation path. First, the fault causal graph of the fault sub-block is constructed as follows: Figure 7 As shown in part (a), the fault in sub-block 1 is transmitted to sub-blocks 2, 3, 5, and 7. The fault variable propagation path can be expressed as Figure 7 As seen in part (b), the variable v 44 The root cause variable causes problems in other variables.

[0098] Table 1. Knowledge-based causal relationships of variables

[0099] Dependent variable Outcome variable Dependent variable Outcome variable Dependent variable Outcome variable <![CDATA[v1]]> <![CDATA[v 44 ]]> <![CDATA[v 14 ]]> <![CDATA[v 15 ,v 16 ,v 18 ]]> <![CDATA[v 44 ]]> <![CDATA[v1]]> <![CDATA[v2]]> <![CDATA[v 42 ]]> <![CDATA[v 15 ]]> <![CDATA[v 49 ]]> <![CDATA[v 45 ]]> <![CDATA[v 15 ,v 16 ,v 18 ]]> <![CDATA[v3]]> <![CDATA[v 43 ]]> <![CDATA[v 16 ]]> <![CDATA[v5]]> <![CDATA[v 46 ]]> <![CDATA[v 47 ]]> <![CDATA[v4]]> <![CDATA[v 45 ]]> <![CDATA[v 17 ]]> <![CDATA[v 19 ]]> <![CDATA[v 47 ]]> <![CDATA[v 10 ]]> <![CDATA[v5]]> <![CDATA[v6]]> <![CDATA[v 18 ]]> <![CDATA[v 50 ]]> <![CDATA[v 48 ]]> <![CDATA[v 14 ]]> <![CDATA[v6]]> <![CDATA[v7,v8,v9,v 42 ,v 43 ,v 44 ]]> <![CDATA[v 19 ]]> <![CDATA[v 18 ,in 50 ]]> <![CDATA[v 49 ]]> <![CDATA[v 17 ]]> <![CDATA[v9]]> <![CDATA[v 51 ]]> <![CDATA[v 20 ]]> <![CDATA[v5]]> <![CDATA[v 50 ]]> <![CDATA[v 18 ,in 19 ]]> <![CDATA[v 10 ]]> <![CDATA[v 20 ]]> <![CDATA[v 21 ]]> <![CDATA[v9]]> <![CDATA[v 51 ]]> <![CDATA[v9,v 21 ]]> <![CDATA[v 11 ]]> <![CDATA[v 12 ,in 13 ]]> <![CDATA[v 22 ]]> <![CDATA[v 11 ]]> <![CDATA[v 52 ]]> <![CDATA[v 11 ,in 22 ]]> <![CDATA[v 12 ]]> <![CDATA[v 48 ]]> <![CDATA[v 42 ]]> <![CDATA[v2]]> <![CDATA[v 13 ]]> <![CDATA[v 46 ]]> <![CDATA[v 43 ]]> <![CDATA[v3]]>

[0100] Table 2 Fault detection rate in sub-blocks

[0101]

[0102]

[0103] Table 3 Comparison results of three methods

[0104]

Claims

1. A knowledge- and data-driven distributed process monitoring method at the plant level, characterized in that: The following steps are included: (1) Collect the data set of normal operating conditions of the process, that is, the training set, denoted as X m×n , m is the number of variables, n is the number of samples; (2) According to the attributes of the variables in the process flow diagram, the knowledge causal graph G = (C, E) is determined, where C is the node set of the dependent variable and E is the node set of the result variable. The data set is divided into several physically meaningful sub-blocks by knowledge causality and recorded as X = [X1, X2, ..., X q ]; (3) Calculate the normalized mutual information value NMI between variables in the sub-block ij , and set a threshold NMI il , NMI ij >NMI il The variables are further divided into separate sub-blocks denoted as X b Represents the b-th sub-block after division; (4) Using random kernel principal component analysis to determine the threshold of each sub-block is recorded as and (5) For the online detection stage, the online collected data Y m×t Steps (2) to (3) are divided into blocks and recorded as (6) Perform random kernel principal component analysis on each sub-block to obtain the detection statistic and (7) If or A fault is detected if and There is no fault; (8) The detection results of each sub-block are Bayesian fused to obtain the global fault detection statistics and SPE GFI , the significance level α is recorded as the global detection control limit; if or SPE GFI >α will be displayed as a fault. And SPE GFI ≤α is normal state; (9) For the detected fault, the contribution rate between sub-blocks is calculated first and recorded as CBCI b ,The sub-block with high contribution rate is defined as the faulty sub-block and recorded as FBS; (10) Perform variable contribution analysis on the fault sub-block, denoted as CVCI b , the variable with high contribution rate is defined as fault variable and recorded as FVS; (11) Use the knowledge causal graph to perform causal analysis on FBS and FVS to obtain the root cause and propagation path of the fault.

2. The method for plant-level distributed process monitoring based on knowledge and data dual drive according to claim 1 is characterized in that: The step (2) is: knowledge causal graph G = (C, E), where C = {c1, c2, ..., c g } and E={e1,e2,…,e f } distribution represents the node set of dependent variables and outcome variables, G∈R g×f It is a knowledge causal matrix, where both rows and columns are represented as nodes; the elements in the matrix are composed of 0 and 1, where 1 represents a causal relationship between two variables and 0 represents no causal relationship.

3. The method for plant-level distributed process monitoring based on knowledge and data dual drive according to claim 1 is characterized in that: The step (3) is as follows: mutual information is to measure the correlation between variables by using the entropy relationship between two variables. The greater the correlation between the variables, the greater the value of mutual information. Mutual information is measured by the difference between entropy and conditional entropy, which is expressed as: I(X;Y)=H(X)-H(X|Y) Normalized mutual information means limiting the value of mutual information to between 0 and 1, which is expressed as: The threshold is set empirically as follows:

4. The method for plant-level distributed process monitoring based on knowledge and data dual drive according to claim 1 is characterized in that: The step (4) is as follows: random kernel principal component analysis is to project the high-dimensional kernel matrix into a low-rank approximate random subspace, to ensure that the eigenvalues ​​and eigenvectors are approximately unchanged to the greatest extent; for the b-th sub-block Use radial basis kernel function to calculate kernel matrix K b : The centralization of the kernel matrix is ​​recorded as: The elements are Before calculating the threshold, you need to implement the random projection algorithm: set a random matrix Ω b ∈R n×l , l≤n, calculate Y b Perform QR decomposition to obtain Q b , Q b satisfy ε is the approximation error, The matrix Limited to B b In this subspace, for B b Perform feature decomposition Among them L b The diagonal is the eigenvalue Matrix sorted from large to small; by Q b U b get Is an approximate eigenvector matrix containing l principal eigenvalues, for the eigenvector Normalize to: Determine the nonlinear principal component of the training data as For the principal element space p b The determination of is determined by the cumulative variance percentage; the detection threshold with a significance level α is expressed as: in, It is p b and Np b F distribution composed of degrees of freedom; It is h b Chi-square distribution with degrees of freedom, g b Is a weight parameter; assuming that the training sample The mean and variance of b and b b ,but 5. The method for plant-level distributed process monitoring based on knowledge and data dual drive according to claim 1 is characterized in that: The step (6) is: for the b-th sub-block The kernel matrix K 1b The radial basis kernel function is used as The centralization of the kernel matrix is ​​recorded as: The elements are Determine the nonlinear principal component of online data as: The test statistic is recorded as: in, Λ b Front p b A diagonal matrix consisting of eigenvalues.

6. The method for plant-level distributed process monitoring based on knowledge and data dual drive according to claim 1, characterized in that: The step (8) is as follows: the Bayesian fusion strategy includes two steps: probability conversion and probability fusion; for the b-th sub-block, firstly, the sub-block monitoring statistics are converted into the sub-block failure probability: in and The calculation is the same as that of the prior probability, which represents the marginal probability, N and F represent normal and abnormal situations; The likelihood probability is expressed as: and and and The calculation is the same; then the sub-block failure probabilities are fused to obtain the global monitoring statistics, which is specifically recorded as: The significance level α is recorded as the global detection control limit; if or SPE GFI >α will be displayed as a fault. And SPE GFI ≤α is the normal state.

7. The method of plant-level distributed process monitoring based on knowledge and data dual drive according to claim 1 is characterized in that: The step (9) is: analyzing the fault sub-blocks of the detected fault, assuming that n F fault samples, the principal component space contribution index TBCI of the b-th sub-block b and residual space contribution index QBCI b Denoted as: when When SPE bi >SPE bα ; Define the combined sub-block contribution index CBCI b Denoted as: The sub-blocks with larger contribution rates are located as the faulty block set.

8. The method for plant-level distributed process monitoring based on knowledge and data dual drive according to claim 1 is characterized in that: The step (10) is: calculating the variable contribution index of the fault block, and the calculation steps for the b-th fault sub-block are: where j = 1, ..., m b , m b is the number of variables in sub-block b, Θ is a fault experience parameter, I 1n =[1,···,1]∈R 1×n , I nn ∈R n×n ;Will with CVSPE b Combine to obtain CVCI b : The variables with larger contribution rates are defined as the fault variable set.

9. The method for plant-level distributed process monitoring based on knowledge and data dual drive according to claim 1, characterized in that: The step (11) is: combining the FBS and FVS obtained in steps (9) and (10) with the knowledge causal graph, performing causal analysis on them, and identifying the root cause and propagation path of the fault; first, at the sub-block level, identifying the causal relationship between the fault sub-blocks based on the knowledge causal graph; the variable set of sub-block B1 is V1, the variable set of sub-block B2 is V2, and the knowledge causal graph between V1 and V2 is represented by G 12 , G 12 The internal node set is V c , the result node set is represented as V e ; The causal relationship between B1 and B2 is: Where B1→B2 means B1 is the fault cause block. Represents the fault sub-blocks of B1 and B2 that are causally related to each other; the causal relationship between the fault sub-blocks is obtained: G FBS =(C FBS ,E FBS ), select the sub-block that has no node pointing to it as the root sub-block; Secondly, at the variable level, the causal relationship between fault variables is G FVS =(C FVS ,E FVS ), through G FVS The source node is used to identify the root cause of the fault.

Images