Data knowledge joint driven dynamic industrial process early fault separation method

Abstract

The application discloses a data-knowledge joint driving dynamic industrial process early fault separation method, belongs to the cross technical field of industrial automation and artificial intelligence, and is used for industrial process fault separation. The method comprises the following steps: collecting measurement data under normal working conditions and performing standardization, calculating typical variable residuals, adopting a sliding window to obtain statistics, and determining the control limit of a detection index; collecting measurement data under abnormal working conditions and performing standardization and smoothing processing, and acquiring a fault direction by singular value decomposition; collecting measurement data under real-time working conditions as test data and performing standardization; calculating typical variable residuals of the test data for fault detection; and based on an optimization criterion, calculating a reconstruction contribution value of a fault sample under different fault directions to determine an actual fault direction, so that fault separation is realized. The method fully utilizes fault direction information, more accurately describes a fault, improves the performance of an early fault separation method, and simultaneously realizes accurate separation of sensor faults and process faults.

Description

Technical Field

[0001] This invention discloses a dynamic industrial process early fault separation method driven by data knowledge, which belongs to the interdisciplinary field of industrial automation and artificial intelligence. Background Technology

[0002] In recent years, numerous sensors have been applied to industrial processes to collect process data in real time and obtain information on process operation status. Thanks to the vast amount of industrial process data, data-driven methods have made significant progress in industrial process monitoring and have attracted widespread attention from both academia and industry. Dynamic industrial processes, due to the inherent time correlation of variables (meaning that current measurements are influenced by past measurements), render traditional static monitoring methods unsuitable for dynamic industrial process monitoring. To effectively address the problem of dynamic process monitoring, some research has proposed improved data-driven methods for fault detection and isolation in dynamic processes. On the other hand, in the early stages of a fault, its amplitude is small and easily masked by normal changes or noise in the process data, making timely detection and correct isolation difficult. As time progresses, the fault amplitude gradually increases, and once it evolves into a serious fault, it will affect the operation of the production process, leading to economic losses. Therefore, early fault detection and isolation are of great significance. Currently, some work has focused on early fault detection in dynamic industrial processes and made some progress, but research methods for early fault isolation are relatively scarce.

[0003] Existing dynamic process fault detection and isolation methods make limited use of process knowledge. In fact, fully utilizing process knowledge in industrial process monitoring could significantly improve fault isolation performance. Fault-related knowledge, especially fault direction reflecting crucial fault information, can be obtained from large historical datasets. How to efficiently utilize this knowledge to enhance fault isolation performance has significant theoretical and practical value.

[0004] Therefore, there is an urgent need to propose a dynamic industrial process early fault separation method driven by data and knowledge to achieve accurate separation of early faults in industrial processes. Summary of the Invention

[0005] The purpose of this invention is to provide a data-knowledge-driven method for early fault separation in dynamic industrial processes, in order to solve the problems of insufficient sensitivity in detecting early minor faults and difficulty in accurately separating and locating fault sources in existing fault diagnosis methods in dynamic, multivariate coupled industrial processes.

[0006] Data- and knowledge-driven methods for early fault isolation in dynamic industrial processes include:

[0007] S1. Collect sensor measurement data under normal operating conditions of industrial process as training dataset, and standardize the training dataset to obtain standardized measurement samples.

[0008] S2. Given the time delay number, construct the Hankel matrix in combination with the measurement sample, calculate the residuals of the typical variables of the measurement sample based on the Hankel matrix, set the sliding window width to calculate the statistics of the residuals of the typical variables, form a statistical data matrix, and calculate the control limit of the Mahalanobis distance detection index based on the statistical data matrix.

[0009] S3. Collect sensor measurement data under abnormal operating conditions in the industrial process as a fault dataset and perform standardization processing to obtain fault data samples. Use the smoothed fault data samples to form a data matrix. Use the singular value decomposition method to obtain the matrix composed of singular vectors of the data matrix. Calculate the Mahalanobis distance detection index of the statistics of the residuals of typical variables in the fault dataset to determine the dimension of the fault direction.

[0010] S4. Collect sensor measurement data under real-time operating conditions of industrial processes as a test dataset and perform standardization processing.

[0011] S5. Calculate the residuals of the typical variables in the test dataset and the statistical vector of the residuals of the typical variables in the test dataset based on the time delay number, the projection matrix of the typical variables, and the diagonal matrix. Calculate the Mahalanobis distance detection index of the test dataset based on the statistical data matrix after zero mean. Determine the occurrence of the fault based on the Mahalanobis distance detection index of the test dataset.

[0012] S6. Construct a fault direction set from all fault directions obtained in step S3. If a fault is detected in step S5, calculate the reconstruction contribution value of the typical variable residuals in the test dataset, and calculate the actual fault direction based on the reconstruction contribution value.

[0013] S1 includes collecting sensor measurement data under normal operating conditions of the industrial process. As a training dataset , For the set of real numbers, For the sample size of the training dataset, For each normal operating condition sample vector, Include One variable, ,right Standardization is performed, which includes calculating the mean and standard deviation of each variable, subtracting the mean from the mean of each variable, and dividing the result by the standard deviation to make the mean of each variable 0 and the variance 1.

[0014] right Standardized data is obtained after standardization processing. , , For standardized measurement samples, This is the transpose symbol.

[0015] S2 includes, S2.1, a given number of past time delays. and future time delay ,set up For the sampling time of the training dataset, ,according to Time sample Construct the corresponding past projection vector and future projection vector :

[0016] ;

[0017] ;

[0018] Will and Arranged in columns, forming the past Hankel matrix. And the future Hankel matrix :

[0019] ;

[0020] ;

[0021] ;

[0022] In the formula, is the number of vectors contained in the Hankel matrix.

[0023] S2 includes S2.2, calculation covariance matrix ,calculate covariance matrix ,calculate and covariance matrix ,based on , and Calculate the whitened cross-covariance matrix :

[0024] ;

[0025] For matrix Perform singular value decomposition:

[0026] ;

[0027] In the formula, and For the reason A matrix composed of singular vectors, It is a left singular vector matrix. It is a right singular vector matrix. For the reason A diagonal matrix composed of the singular values;

[0028] Calculate the projection matrix of future canonical variables Projection matrix of past canonical variables :

[0029] ;

[0030] ;

[0031] calculate Corresponding canonical variable residuals :

[0032] ;

[0033] In the formula, for A column vector of dimension, , The number of residuals for the canonical variables. for The first r lines, for The first r lines, For the reason A diagonal matrix consisting of the first r rows and r columns.

[0034] S2 includes, S2.3, where the width of the sliding window is... Calculate each sliding window The mean vector constitutes the mean statistic. , , for The Middle The mean of each variable; Arranged by row, forming a statistical data matrix. , have OK List, ;calculate mean , According to the zero mean Calculate the corresponding covariance matrix Calculated based on the kernel density estimation method Control limits of Mahalanobis distance detection index .

[0035] S3 includes, S3.1, collecting sensor measurement data under abnormal operating conditions in the industrial process as a fault dataset, including setting measurement variables. When a fault occurs, the data measured by the sensor becomes the fault dataset. The corresponding fault direction is The measured variables are process parameters that characterize the operating status of the industrial process, monitored by sensors; utilizing... The fault dataset is standardized by taking the mean and standard deviation of the standard values. The fault data sample at time is , The time index for the fault dataset samples;

[0036] S3 includes, in S3.2, smoothing the standardized fault dataset using a sliding window technique. Smoothing includes calculating the mean of the fault data within the sliding window and replacing the fault data in the sliding window with the mean of the fault data within the sliding window; then, after smoothing... The fault data sample at time is recorded as , each sliding window Arranged in columns, they form a standardized and smoothed fault data matrix. ,right Perform singular value decomposition:

[0037] ;

[0038] In the formula, and For the reason A matrix composed of singular vectors, For the reason The diagonal matrix formed by the singular values ​​of the elements will... The non-zero singular values ​​contained therein are arranged in descending order. Select... .

[0039] S3 includes, S3.3, from Start with the first column, select The first column indicates the fault direction, then the columns are arranged along the fault direction. Refactor:

[0040] ;

[0041] In the formula, Reconstructed along the fault direction Real-time fault samples, For matrix The first column, Let be the vector of fault amplitudes to be estimated;

[0042] calculate Mahalanobis distance test index for the statistics of canonical variable residuals and with If a comparison is made, , ;like ,choose The next column indicates the fault direction, until... ,matrix The former All columns were selected as fault directions. , for The former List, Fault direction The dimension of.

[0043] S4 includes collecting sensor measurement data under real-time operating conditions of industrial processes as test data. For the time when the test data is sampled, , The test data at each moment is , and The variables are consistent. ,use The mean and standard deviation of Standardization is performed to obtain the processed data. .

[0044] S5 includes, S5.1, based on step S2, based on and calculate Past projection vector and Future projection vector ,based on , and calculate canonical variable residuals ,based on calculate Statistical vector ,Will minus ,get The statistic vector after zero mean ;

[0045] S5 includes S5.2, based on ,calculate Mahalanobis distance detection index :

[0046] ;

[0047] like This indicates that a fault has been detected.

[0048] S6 includes, S6.1, using all the fault directions obtained in step S3 to form a fault direction set when the measured variable fails. , This is an index indicating the direction of the fault when the measured variable fails. , The number of fault directions is equal to the number of variables. For the first When a measured variable fails, the direction of the fault is determined. If a fault is detected in step S5, the calculation is performed. Along Reconstruction contribution value :

[0049] ;

[0050] ;

[0051] In the formula, For the first The direction of failure in the statistical space of the residuals of typical variables when a measurement variable fails. for and The matrix obtained by arranging columns , For the reason indivual A direction matrix stacked row by row ;

[0052] S6 includes S6.2, based on Determine the actual direction of the fault :

[0053] ;

[0054] ;

[0055] In the formula, To find out To obtain the maximum value The value, In order to make To obtain the maximum value The value, For index The direction of the fault.

[0056] Compared with the prior art, the present invention has the following advantages: by making full use of fault direction information, the present invention can more accurately describe the fault, improve the performance of the early fault separation method, and at the same time achieve accurate separation of sensor faults and process faults. Attached Figure Description

[0057] Figure 1 This is a flowchart of the method of the present invention;

[0058] Figure 2 This is a schematic diagram of early fault detection results from sensors in a dynamic industrial process.

[0059] Figure 3 This is a schematic diagram of the early fault isolation results of sensors in a dynamic industrial process;

[0060] Figure 4 This is a schematic diagram of early-stage fault detection results in a dynamic industrial process;

[0061] Figure 5 This is a schematic diagram of the early-stage fault isolation results in a dynamic industrial process. Detailed Implementation

[0062] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention are described clearly and completely below. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0063] Data- and knowledge-driven methods for early fault isolation in dynamic industrial processes include:

[0064] S1. Collect sensor measurement data under normal operating conditions of industrial process as training dataset, and standardize the training dataset to obtain standardized measurement samples.

[0065] S2. Given the time delay number, construct the Hankel matrix in combination with the measurement sample, calculate the residuals of the typical variables of the measurement sample based on the Hankel matrix, set the sliding window width to calculate the statistics of the residuals of the typical variables, form a statistical data matrix, and calculate the control limit of the Mahalanobis distance detection index based on the statistical data matrix.

[0066] S3. Collect sensor measurement data under abnormal operating conditions in the industrial process as a fault dataset and perform standardization processing to obtain fault data samples. Use the smoothed fault data samples to form a data matrix. Use the singular value decomposition method to obtain the matrix composed of singular vectors of the data matrix. Calculate the Mahalanobis distance detection index of the statistics of the residuals of typical variables in the fault dataset to determine the dimension of the fault direction.

[0067] S4. Collect sensor measurement data under real-time operating conditions of industrial processes as a test dataset and perform standardization processing.

[0068] S5. Calculate the residuals of the typical variables in the test dataset and the statistical vector of the residuals of the typical variables in the test dataset based on the time delay number, the projection matrix of the typical variables, and the diagonal matrix. Calculate the Mahalanobis distance detection index of the test dataset based on the statistical data matrix after zero mean. Determine the occurrence of the fault based on the Mahalanobis distance detection index of the test dataset.

[0069] S6. Construct a fault direction set from all fault directions obtained in step S3. If a fault is detected in step S5, calculate the reconstruction contribution value of the typical variable residuals in the test dataset, and calculate the actual fault direction based on the reconstruction contribution value.

[0070] S1 includes collecting sensor measurement data under normal operating conditions of the industrial process. As a training dataset , For the set of real numbers, For the sample size of the training dataset, For each normal operating condition sample vector, Include One variable, ,right Standardization is performed, which includes calculating the mean and standard deviation of each variable, subtracting the mean from the mean of each variable, and dividing the result by the standard deviation to make the mean of each variable 0 and the variance 1.

[0071] right Standardized data is obtained after standardization processing. , , For standardized measurement samples, This is the transpose symbol.

[0072] S2 includes, S2.1, a given number of past time delays. and future time delay ,set up For the sampling time of the training dataset, ,according to Time sample Construct the corresponding past projection vector and future projection vector :

[0073] ;

[0074] ;

[0075] Will and Arranged in columns, forming the past Hankel matrix. And the future Hankel matrix :

[0076] ;

[0077] ;

[0078] ;

[0079] In the formula, is the number of vectors contained in the Hankel matrix.

[0080] S2 includes S2.2, calculation covariance matrix ,calculate covariance matrix ,calculate and covariance matrix ,based on , and Calculate the whitened cross-covariance matrix :

[0081] ;

[0082] For matrix Perform singular value decomposition:

[0083] ;

[0084] In the formula, and For the reason A matrix composed of singular vectors, It is a left singular vector matrix. It is a right singular vector matrix. For the reason A diagonal matrix composed of the singular values;

[0085] Calculate the projection matrix of future canonical variables Projection matrix of past canonical variables :

[0086] ;

[0087] ;

[0088] calculate Corresponding canonical variable residuals :

[0089] ;

[0090] In the formula, for A column vector of dimension, , The number of residuals for the canonical variables. for The first r lines, for The first r lines, For the reason A diagonal matrix consisting of the first r rows and r columns.

[0091] S2 includes, S2.3, where the width of the sliding window is... Calculate each sliding window The mean vector constitutes the mean statistic. , , for The Middle The mean of each variable; Arranged by row, forming a statistical data matrix. , have OK List, ;calculate mean , According to the zero mean Calculate the corresponding covariance matrix Calculated based on the kernel density estimation method Control limits of Mahalanobis distance detection index .

[0092] S3 includes, S3.1, collecting sensor measurement data under abnormal operating conditions in the industrial process as a fault dataset, including setting measurement variables. When a fault occurs, the data measured by the sensor becomes the fault dataset. The corresponding fault direction is The measured variables are process parameters that characterize the operating status of the industrial process, monitored by sensors; utilizing... The fault dataset is standardized by taking the mean and standard deviation of the standard values. The fault data sample at time is , The time index for the fault dataset samples;

[0093] S3 includes, in S3.2, smoothing the standardized fault dataset using a sliding window technique. Smoothing includes calculating the mean of the fault data within the sliding window and replacing the fault data in the sliding window with the mean of the fault data within the sliding window; then, after smoothing... The fault data sample at time is recorded as , each sliding window Arranged in columns, they form a standardized and smoothed fault data matrix. ,right Perform singular value decomposition:

[0094] ;

[0095] In the formula, and For the reason A matrix composed of singular vectors, For the reason The diagonal matrix formed by the singular values ​​of the elements will... The non-zero singular values ​​contained therein are arranged in descending order. Select... .

[0096] S3 includes, S3.3, from Start with the first column, select The first column indicates the fault direction, then the columns are arranged along the fault direction. Refactor:

[0097] ;

[0098] In the formula, Reconstructed along the fault direction Real-time fault samples, For matrix The first column, Let be the vector of fault amplitudes to be estimated;

[0099] calculate Mahalanobis distance test index for the statistics of canonical variable residuals and with If a comparison is made, , ;like ,choose The next column indicates the fault direction, until... ,matrix The former All columns were selected as fault directions. , for The former List, Fault direction The dimension of.

[0100] S4 includes collecting sensor measurement data under real-time operating conditions of industrial processes as test data. For the time when the test data is sampled, , The test data at each moment is , and The variables are consistent. ,use The mean and standard deviation of Standardization is performed to obtain the processed data. .

[0101] S5 includes, S5.1, based on step S2, based on and calculate Past projection vector and Future projection vector ,based on , and calculate canonical variable residuals ,based on calculate Statistical vector ,Will minus ,get The statistic vector after zero mean ;

[0102] S5 includes S5.2, based on ,calculate Mahalanobis distance detection index :

[0103] ;

[0104] like This indicates that a fault has been detected.

[0105] S6 includes, S6.1, using all the fault directions obtained in step S3 to form a fault direction set when the measured variable fails. , This is an index indicating the direction of the fault when the measured variable fails. , The number of fault directions is equal to the number of variables. For the first When a measured variable fails, the direction of the fault is determined. If a fault is detected in step S5, the calculation is performed. Along Reconstruction contribution value :

[0106] ;

[0107] ;

[0108] In the formula, For the first The direction of failure in the statistical space of the residuals of typical variables when a measurement variable fails. for and The matrix obtained by arranging columns , For the reason indivual A direction matrix stacked row by row ;

[0109] S6 includes S6.2, based on Determine the actual direction of the fault :

[0110] ;

[0111] ;

[0112] In the formula, To find out To obtain the maximum value The value, In order to make To obtain the maximum value The value, For index The direction of the fault.

[0113] The method flow of this invention is as follows: Figure 1As shown, sensor measurement data under normal operating conditions of the industrial process are collected as training datasets. The mean and standard deviation of each measurement variable in the training dataset are then calculated, and the training dataset is standardized. The residuals of the typical variables corresponding to each sample in the standardized training dataset are calculated. Given an appropriate sliding window width, the statistics of the residuals of the typical variables in each sliding window are calculated sequentially to form a statistical data matrix. The mean and covariance matrix of this matrix are calculated, and the control limits of the Mahalanobis distance detection index are obtained. Sensor measurement data under historical abnormal operating conditions of the industrial process are collected as fault datasets. The mean and standard deviation of the variables obtained in step S1 are used to standardize multiple fault datasets. The standardized fault datasets are then smoothed. Singular value decomposition is used to obtain the fault direction of each fault dataset and determine the dimension of the fault direction. Sensor measurement data under real-time operating conditions of the industrial process are collected as test data. This test data includes... The variables included are consistent with those included in the training dataset in step S1. Then, the mean and standard deviation from step S1 are used to standardize the test data. The canonical residuals of the standardized test data are calculated. As the test data is continuously collected online and processed, a set of continuous canonical residuals will be obtained. The statistics of the canonical residuals in the sliding window with the same width as in step S2 are calculated. Then, the Mahalanobis distance detection index of the statistics of the canonical residuals of the test data is calculated using the mean and covariance matrix of the statistics data matrix from step S2. The index is compared with the control limit from step S2 to determine whether a fault has occurred. If step S5 determines that the test data contains a fault at the current time, the reconstruction contribution value of the statistics of the canonical residuals of the test data in step S5 along different fault directions is calculated based on the optimization criterion. The different fault directions are obtained from step S3. The fault direction with the largest reconstruction contribution value is determined as the actual fault direction to achieve fault separation.

[0114] To aid in understanding the invention and to visually demonstrate its effectiveness in separating different types of early faults, the following embodiments are used for illustration. First, training data is generated... One normal sample:

[0115] ;

[0116] ;

[0117] ;

[0118] In the formula, , and These are the input vector, state vector, and output vector of the dynamic process, respectively. and The noise is Gaussian noise with zero mean, and the standard deviations are 0.2 and 0.5, respectively. obey Uniform distribution;

[0119] The other matrix parameters are as follows:

[0120] ;

[0121] ;

[0122] ;

[0123] ;

[0124] ;

[0125] ;

[0126] Let the dimension of the measurement vector be... Then, the mean and standard deviation of each variable in the training dataset are calculated, and the training dataset is standardized. The standardization process includes calculating the mean and standard deviation of each variable. For each variable, the corresponding mean is subtracted first, and then the result is divided by the corresponding standard deviation, so that the mean of each variable is reduced to 0 and the variance is reduced to 1, thus completing the standardization process.

[0127] Calculate and retain the control limits for fault detection indicators. Given the past time delay number... and future time delay Both are equal to 4. Calculate the canonical variable residuals for each sample in the standardized training dataset. Choose the width. The sliding window, currently The residuals of all canonical variables in the time-sliding window can be represented by the following matrix:

[0128] ;

[0129] Then, the mean of each variable in the sliding window is calculated using the following formula, and a vector of mean statistics is constructed. :

[0130] ;

[0131] ;

[0132] In the formula, For variable index, , For the index of data points within the sliding window, , For the first The mean of each variable.

[0133] Get each sliding window Arranged by row, forming a statistical data matrix. Statistical data matrix There are 9894 rows and 4 columns. Next, calculate the matrix. The mean of each variable in the vector is formed by the mean of the variables. For the matrix Perform zero-mean normalization and calculate the matrix using the following formula. covariance matrix :

[0134] ;

[0135] In the formula, This represents the statistical data matrix after zero-mean normalization. Then, the control limits for the Mahalanobis distance detection index are calculated using the kernel density estimation method. .

[0136] Fault data was constructed, and related fault directions were obtained. Seven fault datasets were generated using simulation techniques, each consisting of 6000 fault samples, representing the fault occurrence of one measured variable. Each fault dataset was standardized. A sliding window with a width of 100 was used to smooth each fault dataset, and singular value decomposition was performed on the smoothed datasets. Then, the obtained matrix... Starting with the first column, we select the fault direction, reconstruct the fault data, and calculate the Mahalanobis distance detection index of the corresponding canonical variable residual statistics. The dimension of the fault direction is equal to the matrix that makes the reconstructed Mahalanobis distance detection index fall back within the control limits. The minimum dimension.

[0137] Construct test data. Consider two fault types. Fault 1 is a sensor deviation fault, specifically a variable... Fault 1 is a process deviation fault; Fault 2 is a process deviation fault, specifically a variable... A deviation fault occurred. Based on these two scenarios, two sets of test data were generated, each containing 5000 samples, with the fault occurring starting from the 2001st sample.

[0138] Given past time delays and future time delay Both equal 4, which can be obtained from the following two equations. Past projection vector and Future projection vector :

[0139] ;

[0140] ;

[0141] Based on S2 , and Calculate according to the following formula canonical variable residuals :

[0142] ;

[0143] Similarly, using a sliding window with a width of 100, the corresponding window matrix is ​​obtained. :

[0144] ;

[0145] Then, calculate the matrix. The mean of each variable is calculated and a vector of mean statistics is constructed. .Will minus ,get Vector after zero mean Next, based on the covariance matrix obtained in S2... The Mahalanobis distance detection statistic is calculated according to the following formula:

[0146] ;

[0147] and If a comparison is made, If so, it indicates that a fault has been detected.

[0148] Figure 2 and Figure 4 The detection results of the method of this invention for two different faults are shown respectively. It can be seen that after the fault occurs, Significant changes occurred, exceeding the control limits (red line in the figure). The false alarm rates (FAR) for the two faults were 0.00% and 0.70%, respectively, both less than the significance level of 1%, falling within a reasonable range. The fault detection rates (FDR) for this fault detection index were 97.86% and 94.09%, respectively, which are very satisfactory results.

[0149] Early faults are separated and formed into a fault direction set. For each fault sample, the reconstruction contribution value under different fault directions is calculated, and then a judgment is made. The fault direction with the largest reconstruction contribution value is considered to be the direction of the actual fault. Figure 3 and Figure 5The figure illustrates the separation results of the method of this invention for two types of faults, using one test sample as an example for each. The horizontal axis in the figure represents all candidate fault directions, and the vertical axis represents the reconstruction contribution value of different fault directions. Figure 3 The results shown are for the separation of fault 1, using the 2110th sample as an example. Figure 5 The results shown are for fault 2, using the 2930th sample as an example. Using the correct separation rate as the evaluation index for fault separation performance, the correct separation rates of the proposed method are 99.37% and 91.50%, respectively, which are very satisfactory.

[0150] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for dynamic industrial process early fault isolation driven by data and knowledge, characterized in that, The method comprises the following steps: S1, collecting sensor measurement data under normal working conditions of the industrial process as a training data set, and performing standardization processing on the training data set to obtain measurement samples after standardization processing; S2, a time delay number is given, a Hankel matrix is constructed in combination with the measurement samples, typical variable residuals of the measurement samples are calculated based on the Hankel matrix, a sliding window width is set to calculate statistics of the typical variable residuals, a statistics data matrix is constructed, and a control limit of a Mahalanobis distance detection index is calculated based on the statistics data matrix; S3, collecting sensor measurement data under abnormal working conditions of the industrial process as a fault data set and performing standardization processing to obtain fault data samples, using the smoothed fault data samples to construct a data matrix, using a singular value decomposition method to obtain a matrix composed of singular vectors of the data matrix, and calculating a Mahalanobis distance detection index of statistics of typical variable residuals of the fault data set to determine the dimension of the fault direction; S4, collecting sensor measurement data under real-time working conditions of the industrial process as a test data set and performing standardization processing; S5, calculating typical variable residuals of the test data set and a statistics vector of the typical variable residuals of the test data set based on the time delay number, the typical variable projection matrix and the diagonal matrix, calculating a Mahalanobis distance detection index of the test data set based on the zero-mean statistics data matrix, and judging the occurrence of a fault based on the Mahalanobis distance detection index of the test data set; S6, constructing a fault direction set by all the fault directions obtained in step S3, if the occurrence of a fault is detected in step S5, calculating a reconstruction contribution value of the typical variable residuals of the test data set, and calculating an actual fault direction based on the reconstruction contribution value. S2 comprises, S2.1, given past time lag number and future time lag number , set as the sampling time of the training data set, , according to time sample corresponding past projection vector and future projection vector : ； ； will be described below. and arranged in columns, constituting past Hankel matrices and future Hankel matrices : ； ； ； In the formula, is the number of vectors contained in the Hankel matrix; S2 comprises, S2.2, computing the covariance matrix of the covariance matrix of and the covariance matrix of , and the whitened cross-covariance matrix : ； On a matrix singular value decomposition: ； wherein and is a matrix consisting of the singular vectors of is a left singular vector matrix, is a right singular vector matrix, is a diagonal matrix consisting of the singular values of ​​ computing future canonical variable projection matrix and past canonical variable projection matrix : ； ； Computations Corresponding typical variable residuals : ； wherein is a column vector, , is the number of typical variable residuals, is the first r rows of is the first r rows of is a diagonal matrix composed of the first r rows and r columns of . S2 includes, S2.3, setting the sliding window width as , calculating the mean vector of each sliding window , forming the mean statistics , , is the mean of the first variable in ; arranging by rows to form the statistics data matrix , has rows columns, ; calculating the mean , , calculating the corresponding covariance matrix from the zero-meaned , calculating the control limit of the Mahalanobis distance detection index of according to the kernel density estimation method ;​ S3 comprises, S3.1, collecting sensor measurement data under abnormal working conditions of the industrial process as a fault data set, including measurement variables The sensor measurement data is a fault data set The corresponding fault direction is The measurement variables are process parameters monitored by sensors in the industrial process, representing the running state of the industrial process; the fault data set is standardized by using the mean and standard deviation The standardized fault data set The fault data sample at time t is , The time index of the fault data set sample S3 includes, S3.2, smoothing the standardized failure data set by using sliding window technique, the smoothing includes calculating the mean value of the failure data in the sliding window, replacing the failure data in the sliding window with the mean value of the failure data in the sliding window; the standardized failure data matrix after smoothing is obtained by arranging the mean value of each sliding window in rows and columns The failure data sample at time t is denoted as The mean value of each sliding window is denoted as Arranging the mean value of each sliding window in rows and columns, a failure data matrix after standardization and smoothing is obtained The failure data matrix after standardization and smoothing is denoted as singular value decomposition is performed on the failure data matrix after standardization and smoothing ； In the formula, and For the reason A matrix composed of singular vectors, For the reason The diagonal matrix formed by the singular values ​​of will The non-zero singular values ​​contained therein are arranged in descending order. Select... ; S3 comprises, S3.3, reconstructing starting with the first column of the first column of the failure direction, and then along the failure direction reconstructing ； In the formula, is the reconstructed fault direction along the fault is the time fault sample, is the first column of the matrix is the first column of the matrix is the fault amplitude vector to be estimated; Computing Mahalanobis distance detection index of the statistics of typical variable residuals , and compared with , if , ; if , the next column of is selected as the fault direction, until , the first columns of matrix are all selected as the fault direction, , are the first columns of , and is the dimension of the fault direction . S5 includes, S5.1, based on step S2, based on and calculate Past projection vector and Future projection vector ,based on , and calculate canonical variable residuals ,based on calculate Statistical vector ,Will minus ,get The statistic vector after zero mean ; S5 comprises, S5.2, based on , calculating the Mahalanobis distance detection indicator : ； If , it is indicated that a fault has been detected; S6 includes, S6.1, using all the fault directions obtained in step S3 to form a fault direction set when the measured variable fails. , This is an index indicating the direction of the fault when the measured variable fails. , The number of fault directions is equal to the number of variables. For the first When a measured variable fails, the direction of the fault is determined. If a fault is detected in step S5, the calculation is performed. Along Reconstruction contribution value : ； ； wherein is the statistical space of the fault direction of the typical variable residual when the th is the th matrix obtained by arranging the columns, , is the direction matrix stacked by rows, ;​ S6 comprises, S6.2, determining, based on determining the actual fault direction : ； ； wherein To find the values of that maximize , To find the values of that maximize , is the failure direction indexed by .

2. The data-knowledge co-driven dynamic industrial process early fault isolation method according to claim 1, characterized in that, S1 includes collecting sensor measurement data under normal operating conditions of the industrial process. As a training dataset , For the set of real numbers, For the sample size of the training dataset, For each normal operating condition sample vector, Include One variable, ,right Standardization is performed, which includes calculating the mean and standard deviation of each variable, subtracting the mean from the mean of each variable, and dividing the result by the standard deviation to make the mean of each variable 0 and the variance 1. To standardized data , , measurement sample after standardization, is a transpose symbol.

3. The data-knowledge co-driven dynamic industrial process early fault isolation method according to claim 2, characterized in that, S4 includes collecting sensor measurement data in real-time working condition of the industrial process as test data, For the sampling time of test data, , The test data at the time is , Consistent with the variable of , , using the mean and standard deviation of Standardize To get the processed data .

Images