A non-Gaussian industrial process fault diagnosis method based on multi-decomposition kernel entropy learning

Abstract

The present invention discloses a non-Gaussian industrial process fault diagnosis method based on multi-decomposition kernel entropy learning, comprising: using information entropy to project process data in non-Gaussian industrial data into a high-dimensional space to obtain a process data matrix G in the high-dimensional space; separating a redundant information matrix G from the process data matrix G R ; Use singular value decomposition technology to decompose the redundant information matrix G R Decompose the load matrix to obtain a projection matrix; use the projection matrix to obtain the redundant information matrix G R Decompose the fault-insensitive part G N , remove the fault-insensitive part G from the process data matrix G N , get the process data matrix G Y1 , build a data model; build a fault detection strategy; combine detection indicators to build new fault diagnosis indicators. The method of the present invention divides the process data to the greatest extent through two decompositions, which can greatly reduce the false alarm rate of faults. For faults in non-Gaussian industrial processes, this method can restore the original information of the data and determine the fault variables more accurately.

Description

Technical Field

[0001] The present invention belongs to the technical field of fault diagnosis, and in particular relates to a non-Gaussian industrial process fault diagnosis method based on multi-decomposition kernel entropy learning. Background Art

[0002] In recent years, data-driven fault diagnosis methods have been widely used in complex industrial processes such as chemical, pharmaceutical and metallurgical manufacturing, and have achieved remarkable results. However, most existing methods consider the nonlinear characteristics of complex industrial processes, directly use the collected offline data to establish data models, and monitor the system operation status in real time based on relevant statistical characteristics. However, these methods often ignore the non-Gaussian characteristics existing in the actual operation process, and have certain limitations in real-time and accuracy. If the existing data-driven fault diagnosis methods are directly applied to the diagnosis process of actual industrial processes, the non-Gaussian characteristics existing in the operation process will affect the effectiveness of the data model, resulting in underreporting and false alarms in the fault diagnosis results, causing unnecessary shutdowns, maintenance and overhauls, and thus causing certain economic losses to the enterprise.

[0003] At the same time, the existing non-Gaussian fault diagnosis methods use kernel learning methods to directly establish mathematical models, which fail to give the correct relationship between process variables and faults. For example, when a component of an industrial system fails, the staff will check and repair related faults based on the fault information and data model. Inaccurate data models will cause certain deviations in the diagnosis results, extend the maintenance time, increase maintenance costs, and affect the maintenance and maintenance of industrial processes. Although some methods have further reconstructed the data model, they have failed to give a good correlation between variables and faults. The fundamental reason for the above results is that too much irrelevant or redundant system information is retained in the data modeling process, and the system variables affected by the fault cannot be accurately locked, which affects the accuracy of the data model.

[0004] At the same time, when a system fails, existing fault methods usually analyze the fault information to locate the fault variable and give the fault troubleshooting results. However, when the kernel learning method performs dimensionality increase processing on nonlinear non-Gaussian data, it will destroy the original data information in the data, resulting in the staff being unable to troubleshoot the fault variable based on the fault information and data model, affecting the actual application effect of methods such as contribution graphs. In addition, the existing nonlinear fault diagnosis methods only integrate the second-order statistical indicators such as variance and covariance of the data to give the fault location results. However, the assumptions of the second-order statistics are too conservative, ignoring the key information under other statistics, which will further destroy the relationship between the fault variables, resulting in inaccurate fault diagnosis results that cannot meet the production conditions of actual industrial systems. Therefore, in the fault diagnosis stage, it is very necessary to effectively use non-Gaussian high-order information to obtain accurate fault results. Summary of the invention

[0005] The technical problem to be solved by the present invention is how to effectively utilize non-Gaussian high-order information to obtain accurate fault results; the present invention proposes a non-Gaussian industrial process fault diagnosis method based on multi-decomposition kernel entropy learning, which uses the high-order information of the data to establish a model, avoiding the problem of missing information in the data modeling process used in existing methods, and can provide accurate fault diagnosis results by optimizing the data through multiple decompositions.

[0006] A non-Gaussian industrial process fault diagnosis method based on multi-decomposition kernel entropy learning of the present invention comprises the following steps:

[0007] Step 1, collecting operation data of a non-Gaussian industrial system, wherein the operation data includes process data and product quality data;

[0008] Step 2: Use information entropy to project the process data in non-Gaussian industrial data into a high-dimensional space to obtain a process data matrix in the high-dimensional space.

[0009] Step 3: Build an initial data model for the process data matrix The relationship matrix calculation process between the quality data matrix Y and the process data matrix is ​​corrected using orthogonal filtering technology Removal of process data matrix The redundant information matrix in

[0010] Step 4: Use singular value decomposition technology to decompose the redundant information matrix Decompose the load matrix to obtain the projection matrix;

[0011] Step 6: Use the projection matrix to extract the redundant information matrix Decompose the parts that are not sensitive to faults From the process data matrix Remove the insensitive parts to the fault Get the process data matrix The final data model is constructed to accurately detect fault conditions:

[0012] Step 7: Reconstruct the fault detection strategy;

[0013] Step 8: Combine the detection indicators and construct new fault diagnosis indicators;

[0014] Furthermore, the process data matrix in step 2 The solution process is as follows:

[0015] Information entropy is used to project non-Gaussian industrial process data into high-dimensional space;

[0016]

[0017] Where λ represents the order of the function, and p represents the probability density of the function. Z(p) is the high-order information entropy function of the process data U.

[0018] Probability density function of process data U Calculated as follows:

[0019]

[0020] in, is the probability density function, is the high-dimensional expression of the process data U, and the probability density function is The solution is that σ is the smoothing kernel parameter.

[0021] use Process Data Matrix Estimation is performed to provide a high-dimensional data space for processing non-Gaussian nonlinear features in process data;

[0022]

[0023] in, is the normalized new matrix of the process data matrix G in high-dimensional space, 1 N is a matrix whose values ​​are all 1.

[0024] By solving the above formula, we can get the normalized high-dimensional space process data matrix:

[0025] Furthermore, in step 3, the initial data model is constructed as follows:

[0026]

[0027] Among them, Y represents the product quality matrix, B is the relationship matrix, which is solved by the principal component regression method.

[0028] Furthermore, in step 3, orthogonal filtering technology is used to extract the process data matrix Separate the redundant information matrix The process is as follows:

[0029] First, the principal component analysis algorithm is performed on the product quality matrix Y to obtain:

[0030] T=YQ

[0031] Where T is the principal matrix, Q is the load matrix of Y;

[0032] Then the relationship matrix B is processed by the truncation method to determine the number of principal elements A;

[0033] According to the number of main elements A, the intermediate matrix M is calculated in a loop i :

[0034]

[0035] in, represents the process data matrix in the i-th cycle, T i represents the pivot matrix in the i-th cycle.

[0036] For the intermediate matrix M i Perform eigenvalue decomposition to obtain:

[0037]

[0038] Among them, v i represents the eigenvector matrix in the i-th cycle; n represents the number of samples; retains non-zero values ​​λ i The corresponding feature vectors are sorted.

[0039] Specifically, the process data matrix is ​​extracted from the quality matrix Y The associated pivot matrix T Y :

[0040]

[0041] Using its main element matrix part T Y Update a high-dimensional matrix of process data:

[0042]

[0043] in, express The new matrix obtained after the i+1th cycle;

[0044] The initial value of i is 1, and the above operation is repeated until i = A, and the process data matrix is ​​obtained. The process data matrix This is the matrix after removing redundant information;

[0045] The redundant information matrix is ​​written as:

[0046]

[0047] Furthermore, in step 4, the redundant information matrix is ​​decomposed using singular value decomposition technology. Decompose the load matrix to obtain the projection matrix, as follows:

[0048] The objective function is constructed as follows:

[0049] maxcov(P R ,Q)

[0050] sP R *V=I A

[0051] Among them, R is the conversion matrix, Q is the load matrix of Y, P R for The load matrix.

[0052] The load matrix P is analyzed by using singular value decomposition technology. R To break it down:

[0053] P R =Γ T Δ P D T

[0054] Among them, Γ T =[Γ Y Γ N ], which represents the projection matrix.

[0055] Furthermore, in step 5, the projection matrix is ​​used to obtain the redundant information matrix Decompose the parts that are not sensitive to failure And from the process data matrix Remove the insensitive parts to the fault Rebuild the data model. The specific steps are as follows:

[0056] Parts not sensitive to failure It is expressed as follows:

[0057]

[0058] Process Data Matrix Re-expressed as follows:

[0059]

[0060] in,

[0061] Reconstruct the data model as follows:

[0062]

[0063] Among them, B y is the new relationship matrix, which is solved by the principal component regression algorithm.

[0064] Furthermore, the fault detection strategy is constructed in step 6 as follows:

[0065] When the process data matrix of CS discrete statistics is used When an alarm occurs partially, it means that a fault that affects the non-Gaussian industrial system has occurred; otherwise, no fault has occurred.

[0066] Furthermore, constructing the fault detection strategy in step 6 also includes:

[0067] When T 2 Statistical When the matrix alarm occurs, it means that the fluctuations occurring in the system are normal.

[0068] Furthermore, the combined detection index in step 7 refers to the combined CS discrete statistics and T 2 Statistics;

[0069] For normal process data samples Merged statistic Ψ N It is expressed as:

[0070]

[0071] For the fault process data sample Merged statistic Ψ F It is expressed as:

[0072]

[0073] Construct a new diagnostic indicator Tr, which is expressed as follows:

[0074]

[0075] in, is the combined statistic of the normal process data sample Ψ N The column mean of is the combined statistic of the fault process data samples Ψ F The column mean of is the normal sample pooling statistic Ψ N The standard deviation of .

[0076] Furthermore, the non-Gaussian industrial process fault diagnosis method based on multi-decomposition kernel entropy learning of the present invention also includes an online stage:

[0077] Step 1: For the detection sample u online , by using information entropy to transform it into G online , and calculate

[0078]

[0079] Step 2: Use the projection matrix to process the data G online,R To decompose;

[0080] G online,N =G online,R Γ N Γ N T

[0081]

[0082] Step 3: Calculate the CS statistics and T of the online samples according to the fault detection strategy. 2 Statistics, detect the current state of the non-Gaussian industrial process. When the CS discrete statistics issues an alarm, it can be determined that a fault has occurred in the process and enter step 4;

[0083] Step 4: Combine the detection index CS and Obtain a new diagnostic indicator Tr and restore u online ; Step 5, using the original process data, calculate the value of the fault diagnosis index Tr; when the value of Tr is greater than the set threshold, it is determined that the sample is the main cause of the fault.

[0084] Beneficial effects:

[0085] The present invention provides a quality-related multi-decomposition kernel entropy fault diagnosis method, especially a data modeling and fault diagnosis method for non-Gaussian industrial processes. The method uses high-order information entropy and kernel learning methods to process non-Gaussian features in the data and construct a preliminary high-order process data matrix; further, by decomposing the process data twice, reconstructing the data model and designing fault detection statistics and fault diagnosis strategies to ensure its sensitivity to faults; on this basis, by fusing fault detection indicators and restoring the original information of data samples, the fault variables can be determined more accurately and the false alarm rate can be reduced.

[0086] The present invention proposes a kernel learning method based on multiple decompositions to solve the problems of non-Gaussian characteristics and false alarms in data-driven fault diagnosis. (1) This method can effectively process non-Gaussian data multiple times and retain more effective information in data modeling. (2) This method can divide process data to the greatest extent through two decompositions, which can greatly reduce the false alarm rate. (3) This method can restore the original information of the data for faults in non-Gaussian industrial processes and can more accurately determine the fault variables. BRIEF DESCRIPTION OF THE DRAWINGS

[0087] Figure 1 It is a flow chart of the method of the present invention. DETAILED DESCRIPTION

[0088] The present invention establishes a data model for non-Gaussian industrial processes and designs a fault diagnosis scheme to locate fault variables, thereby improving fault diagnosis accuracy and reducing the false alarm rate of faults.

[0089] A non-Gaussian industrial process fault diagnosis method based on multi-decomposition kernel entropy learning of the present invention comprises an offline training phase and an online phase;

[0090] The offline training phase includes the following steps:

[0091] Step 1: Collecting operation data of a non-Gaussian industrial system, wherein the operation data includes process data and product quality data;

[0092] Process data and product quality data are defined as U and Y, where:

[0093] U=[u1 u2…u N ]

[0094] Y=[y1 y2…y N ]

[0095] Wherein, N represents the number of variables. The process data U may be variables such as pressure and temperature of an industrial process, and the product quality data Y may be factors such as thickness, purity and life of a product.

[0096] Step 2: Use information entropy to project non-Gaussian industrial process data into a high-dimensional space;

[0097]

[0098] Where λ represents the order of the function, p represents the probability density of the function, and Z(p) is the high-order information entropy function of U. Probability density function of U It can be calculated as follows:

[0099]

[0100] in, is the probability density function, is a high-dimensional expression of U, which can be obtained by Parzen window technique. Solve and get. σ is the smoothing kernel parameter.

[0101] Considering that the high-order information entropy function Z(p) of U is a monotonic function and difficult to estimate, we can directly use The process data is estimated and a high-dimensional data space is given to handle the non-Gaussian nonlinear characteristics existing in the process data.

[0102]

[0103] in, is the process data matrix in high-dimensional space, 1 N is a matrix whose values ​​are all 1.

[0104] By solving the above formula, we can get the normalized high-dimensional space process data matrix

[0105] Step 3: Construct an initial data model to describe the relationship between the process data matrix G and the quality data matrix, and then use the orthogonal filtering technology to correct the initial data model; construct the following initial data model:

[0106]

[0107] Among them, B is the relationship matrix, which can be solved by the principal component regression method.

[0108] The correction method is as follows:

[0109] First, the principal component analysis algorithm is performed on the product quality matrix Y to obtain:

[0110] T=YQ

[0111] Where T is the principal matrix and Q is the load matrix;

[0112] Then the relationship matrix B is processed by the truncation method to determine the number of principal elements A, and then the normalized new matrix is ​​corrected using the orthogonal filtering technology. Separate the redundant information in the data matrix and build a revised data model.

[0113] According to the number of main elements A, the intermediate matrix M is calculated in a loop i :

[0114]

[0115] in, represents the process data matrix in the i-th cycle, T i represents the pivot matrix in the i-th cycle.

[0116] Then for the intermediate matrix M i Perform eigenvalue decomposition to obtain:

[0117]

[0118] Among them, v i represents the eigenvector matrix in the i-th cycle; n represents the number of samples; this paper retains the non-zero value λ i The corresponding feature vectors are sorted to retain more effective information to the greatest extent possible while removing redundancy.

[0119] Specifically, the process data matrix is ​​extracted from the quality matrix Y The associated pivot matrix T Y :

[0120]

[0121] Further use its main element matrix part T Y Update a high-dimensional matrix of process data:

[0122]

[0123] in, express The new matrix obtained after the i+1th cycle needs to be updated continuously.

[0124] The above operation is repeated until i = A. After the loop ends, remove The new process data matrix can be obtained by using some redundant information in

[0125] The updated data model can be expressed as:

[0126]

[0127] in, Represents the redundant information matrix. When the new process data matrix With the new relationship matrix B new When changes occur, it will affect the product quality matrix Y.

[0128] Step 4: Use singular value decomposition technique to process data The load matrix P R Decompose to obtain the projection matrix; to ensure the accuracy of process data matrix decomposition, the following objective function can be constructed

[0129] maxcov(P R ,Q)

[0130] sP R *V=I A

[0131] Among them, R is the conversion matrix, Q is the load matrix of the quality data matrix Y, P R for The load matrix P is R To break it down:

[0132] P R =Γ T Δ P D T

[0133] Among them, Γ T =[Γ Y Γ N ], which represents the projection matrix.

[0134] Step 5: Use the projection matrix to modify the data model Reconstruct and give the final data model:

[0135] Using the constructed projection matrix Γ Y and Γ N The redundant information matrix Decompose into and Two parts. With the new process data matrix The merge is performed and a new data model is finally given.

[0136]

[0137] in,

[0138] Reconstruct the data model as follows:

[0139]

[0140] High-dimensional representation of process data Can be decomposed into and Two parts, For the part that is not sensitive to faults, Represents the process data matrix that is sensitive to faults. y is a new relationship matrix, which can be solved by principal component regression algorithm. Furthermore, the reconstructed mathematical model can be used to observe the state of non-Gaussian industrial processes;

[0141] Step 6: Build a fault detection strategy;

[0142] Using Cauchy-Schwarz (CS) discrete statistics and T 2 Statistics are used to construct fault detection strategies; when the CS discrete statistics process data matrix is ​​used When an alarm occurs in some cases, it means that a fault that affects the non-Gaussian industrial system has occurred; otherwise, no fault has occurred in the industrial process;

[0143] When T 2 Statistical When the matrix alarm occurs, it means that the fluctuations in the system are normal and will not affect the normal operation of the system.

[0144] Step 7: Combine the detection indicators and construct new fault diagnosis indicators;

[0145] The combined detection index refers to the combined CS discrete statistics and T 2 Statistics;

[0146] For normal process data samples Merged statistic Ψ N It is expressed as:

[0147]

[0148] For the fault process data sample Merged statistic Ψ F It is expressed as:

[0149]

[0150] Construct a new diagnostic indicator Tr, which is expressed as follows:

[0151]

[0152] in, is the combined statistic of the normal process data sample Ψ N The column mean of is the combined statistic of the fault process data samples Ψ F The column mean of is the combined statistic of the normal process data sample Ψ N The standard deviation of

[0153] Online stage:

[0154] Step 1: Process the test sample;

[0155] For the detection sample u online , by using information entropy to transform it into And calculate

[0156]

[0157] Step 2: Decompose the process data using the projection matrix;

[0158] According to the constructed projection matrix To decompose;

[0159]

[0160] Step 3: According to the fault detection strategy, the current state of the non-Gaussian industrial process is detected;

[0161] Calculate CS and T for online samples 2 Statistics and detect the current state of non-Gaussian industrial processes.

[0162]

[0163] in, express The column mean of .

[0164] Using CS discrete statistics Partial design test indicators, and then use T 2 The statistic is the part that is insensitive to the fault Design corresponding statistical indicators. When the CS discrete statistics issue an alarm, it can be judged that a fault has occurred in the process, which has a great impact on product quality.

[0165] Step 4: Combine the detection index CS and Obtain a new diagnostic indicator Tr and restore u online ;

[0166] The diagnostic indicator Tr includes The second-order statistical information of CS also contains the high-order information of CS statistics, which is more general for fault diagnosis results.

[0167] Data online After dimensionalization, it becomes The original information is destroyed and the existing methods cannot be used to directly obtain the fault diagnosis results. It is necessary to apply the nuclear equivalent method to restore it.

[0168] When a fault is detected, statistical indicators are combined and Restore u online The original data information;

[0169]

[0170] Among them, Π is the transformation matrix, which can be obtained by the kernel function equivalent transformation method. Q is the load matrix of the quality data matrix Y.

[0171] Step 5: Using the original data information and diagnostic indicators, determine the fault variable;

[0172] The original process data information and the constructed diagnostic index Tr are used to analyze the current fault variable to determine the fault variable. When the Tr of a sample is greater than the set threshold, an alarm is immediately issued, and it is determined that the process data sample is the main cause of the fault.

[0173] The present invention belongs to a fault diagnosis method for non-Gaussian industrial processes, which can obtain an accurate data model by decomposing the data model multiple times. Specifically, the non-Gaussian nonlinear characteristics existing in the data sample are processed according to the information entropy and kernel learning method to retain the relevant information to the greatest extent; then, the data model is reconstructed by introducing the double decomposition technology to ensure the sensitivity of the data model to the fault intensity and reduce the situation of false fault alarms; finally, the fault detection indicators of the data model are integrated to restore the original information of the data sample to ensure the accuracy of fault diagnosis.

Claims

1. A non-Gaussian industrial process fault diagnosis method based on multi-decomposition kernel entropy learning, characterized in that: The steps include: Step 1, collecting operation data of a non-Gaussian industrial system, wherein the operation data includes process data and product quality data; Step 2: Use information entropy to project the process data in non-Gaussian industrial data into a high-dimensional space to obtain a process data matrix in the high-dimensional space. Step 3: Build an initial data model to describe the process data matrix and the relationship between the quality data matrix Y; The initial data model is constructed as follows: Among them, Y represents the product quality matrix, B is the relationship matrix, which is solved by the principal component regression method; orthogonal filtering technology is used to obtain the process data matrix Separate the redundant information matrix The process is as follows: First, perform the principal component analysis algorithm on the product quality matrix Y to obtain: T=YQ Where T is the principal matrix, Q is the load matrix of Y; Then the relationship matrix B is processed by the truncation method to determine the number of principal elements A; According to the number of main elements A, the intermediate matrix M is calculated in a loop i : in, represents the process data matrix in the i-th cycle, T i represents the main element matrix in the i-th cycle; For the intermediate matrix M i Perform eigenvalue decomposition to obtain: Among them, v i represents the eigenvector matrix in the i-th cycle; n represents the number of samples; retains non-zero values ​​λ i The corresponding feature vectors are sorted; Specifically, extract the process data matrix from the quality matrix Y The associated pivot matrix T Y : Using its main element matrix part T Y Update a high-dimensional matrix of process data: in, express The new matrix obtained after the i+1th cycle; The initial value of i is 1, and the above operation is repeated until i = A, and the process data matrix is ​​obtained. The process data matrix is the matrix after removing redundant information; Redundant Information Matrix Written as: Step 4: Use singular value decomposition technology to decompose the redundant information matrix Decompose the load matrix to obtain the projection matrix, as follows: The objective function is constructed as follows: maxcov ( P R ,Q) s.t.P R *V=I A Among them, R is the conversion matrix, Q is the load matrix of Y, P R for The load matrix; The load matrix P is analyzed by using singular value decomposition technology. R To break it down: P R =C T D P D T Among them, Γ T =[Γ Y Γ N ], represents the projection matrix; Step 5: Use the projection matrix to obtain the redundant information matrix Decompose the insensitive parts to the fault From the process data matrix Remove the parts that are not sensitive to failure Get the process data matrix that is sensitive to faults Rebuild the data model. The specific steps are as follows: Parts not sensitive to failure It is expressed as follows: Process Data Matrix Re-expressed as follows: in, Reconstruct the data model as follows: Among them, B y is the new relationship matrix, which is solved by the principal component regression algorithm; Step 6: Build a fault detection strategy as follows: When the process data matrix of CS discrete statistics is used When some alarms occur, it means that a fault that affects the non-Gaussian industrial system has occurred; otherwise, no fault has occurred; When T 2 Statistical When the matrix alarm occurs, it means that the fluctuations in the system are normal fluctuations; Step 7: Combine the detection indicators and construct new fault diagnosis indicators; The combined detection index refers to the combined CS discrete statistics and T 2 Statistics; For normal process data samples Merged statistic Ψ N It is expressed as: For process data samples when a fault occurs Merged statistic Ψ F It is expressed as: Construct a new diagnostic indicator Tr, which is expressed as follows: in, is the combined statistic of the normal process data sample Ψ N The column mean of is the combined statistic of the fault process data samples Ψ F The column mean of is the combined statistic of the normal process data sample Ψ N The standard deviation of .

2. According to claim 1, a non-Gaussian industrial process fault diagnosis method based on multi-decomposition kernel entropy learning is characterized in that: Process data matrix in step 2 The solution process is as follows: Using information entropy to project non-Gaussian industrial process data into high-dimensional space; Where λ represents the order of the function, p represents the probability density of the function; Z(p) is the high-order information entropy function of the process data U; Probability density function of process data U Calculated as follows: in, is the probability density function, is the high-dimensional expression of the process data U, and the probability density function is The solution is that σ is the smoothing kernel parameter; use Process Data Matrix Estimation is performed to provide a high-dimensional data space for processing non-Gaussian nonlinear features in process data; in, is the normalized new matrix of the process data matrix G in high-dimensional space, 1 N is a matrix whose values ​​are all 1; By solving the above formula, we can get the normalized high-dimensional space process data matrix:

3. According to claim 1, a non-Gaussian industrial process fault diagnosis method based on multi-decomposition kernel entropy learning is characterized in that: Also includes online phase: Step 1: For the detection sample u online , by using information entropy to transform it into G online , and calculate Step 2: Use the projection matrix to process the data G online,R To decompose; Step 3: Calculate the CS statistics and T of the online samples according to the fault detection strategy. 2 Statistics, detect the current state of the non-Gaussian industrial process. When the CS discrete statistics issues an alarm, it can be determined that a fault has occurred in the process and enter step 4; Step 4: Combine the detection index CS and Obtain a new diagnostic indicator Tr and restore u online ; Step 5, using the original process data, calculate the value of the fault diagnosis index Tr; when the value of Tr is greater than the set threshold, it is determined that the sample is the main cause of the fault.

Images