A closed loop process monitoring method based on improved dynamic latent variable analysis

Abstract

The application provides a closed loop process monitoring method based on improved dynamic latent variable analysis, and relates to the technical field of industrial process monitoring, and specifically comprises the following steps: collecting a section of sensor measurement data under normal working conditions of an industrial process as training data; calculating the neighborhood weight matrix of input data and output data respectively; establishing an improved dynamic latent variable analysis model to obtain the weight matrix and load matrix of input and output; constructing the feature matrix of input and output, calculating the covariance matrix and the control limit of the reduced rank Mahalanobis distance index; collecting test data, calculating the feature vector of input and output by using the projection direction matrix, calculating the reduced rank Mahalanobis distance index, and comparing with the control limit to judge whether a fault occurs or not. The technical scheme of the application overcomes the problem in the prior art that the detection method based on the open loop assumption design is difficult to effectively distinguish process faults, interference changes and normal adjustment behaviors of the controller, thereby causing the problems of missed report or false positive rate increase.

Description

Technical Field

[0001] This invention relates to the field of industrial process monitoring, and more specifically to a closed-loop process monitoring method based on improved dynamic latent variable analysis. Background Technology

[0002] In recent years, with the increasing complexity and scale of industrial processes, the safety and reliability of system operation have become a focus of attention. Data-driven fault detection methods have gained widespread attention in the field of process monitoring because they do not require precise mechanistic models and can extract system characteristics based solely on historical process data. Among them, methods based on dynamic latent variable analysis effectively capture the dynamic characteristics of industrial processes by constructing a low-dimensional latent variable space that can characterize the temporal correlation of processes. Compared with traditional static methods, they have shown significant advantages in fault detection performance and have become an important tool for dynamic process monitoring.

[0003] In actual industrial processes, to ensure stable operation and product quality, most industrial processes employ closed-loop control structures. The introduction of closed-loop systems significantly enhances the system's ability to track setpoints and suppress external disturbances, but it also presents new challenges for fault detection. Specifically, under closed-loop control, the controller actively adjusts the input to compensate for deviations between the process output and the setpoint. This can lead to fault information being masked or distorted by the control action, making it difficult for detection methods based on open-loop assumptions to effectively distinguish between process faults, disturbance changes, and normal controller behavior, resulting in increased false alarm or missed alarm rates. Furthermore, the propagation path of faults in closed-loop systems is more complex, and the coupling between different loops further increases the difficulty of fault feature extraction. Although a few methods have addressed the challenges of closed-loop process monitoring, methods based on dynamic latent variable analysis are even rarer.

[0004] Therefore, there is a need for a closed-loop process monitoring method based on improved dynamic latent variable analysis that can achieve fault monitoring of multiple closed-loop processes without requiring an accurate analytical model. Summary of the Invention

[0005] The main objective of this invention is to provide a closed-loop process monitoring method based on improved dynamic latent variable analysis, in order to solve the problem that existing detection methods based on the open-loop assumption are difficult to effectively distinguish between process faults, disturbance changes and normal control behavior of the controller, resulting in a higher rate of missed or false alarms.

[0006] To achieve the above objectives, this invention provides a closed-loop process monitoring method based on improved dynamic latent variable analysis, specifically including the following steps: S1 collects a segment of sensor measurement data under normal operating conditions in the industrial process, and divides the sensor measurement data into input data and output data as training data.

[0007] S2 standardizes the input and output data respectively, and uses the k-nearest neighbor method to calculate the neighborhood weight matrix of the input and output data respectively.

[0008] S3. Given the number of latent variables, establish an improved dynamic latent variable analysis model to obtain the input and output weight matrices and loading matrices.

[0009] S4. Calculate the projection direction matrix using the weight matrix and load matrix from step S3, construct the input and output feature matrices, standardize the feature matrices, and calculate the control limits of the covariance matrix and the reduced-rank Mahalanobis distance index.

[0010] S5. Collect sensor measurement data under real-time operating conditions of the industrial process. Divide the sensor measurement data under real-time operating conditions into input and output data as test data. The measurement variables in the test data correspond to the measurement variables in the training data in step S1.

[0011] S6. The test data is standardized. The input and output feature vectors are calculated using the projection direction matrix in step S4. The feature vectors are standardized, and the reduced-rank Mahalanobis distance index is calculated. The index is compared with the control limit in step S4 to determine whether a fault has occurred.

[0012] Furthermore, step S1 specifically includes: Given the dynamic order of the input data and the dynamic order of the output data Calculate the dynamic order of the industrial process system that needs to be modeled. ,collect The sample data at each time point is divided into input data and output data based on whether the variable is a process input variable or a process output variable. This is a function to find the maximum value.

[0013] Furthermore, step S2 specifically includes the following steps: S2.1, For the input data matrix, calculate the mean and standard deviation of each variable. Subtract the mean from the standard deviation of each variable, and divide the result by the standard deviation to make the mean of each variable 0 and the variance 1. This standardized matrix is ​​denoted as... After standardizing the output data matrix, it is denoted as... ;in, For the set of real numbers, The number of samples in the training data. The number of input variables. It is the first The input samples at each time point, where , The number of variables in the output. It is the first The output samples at each time step, where .

[0014] S2.2, using the k-nearest neighbor method to calculate The neighborhood weight matrix, given the number of neighborhoods is Calculate the Euclidean distance between each input sample, and find the nearest neighbor for each input sample. Each sample is taken as a nearest neighbor of the input sample, and a nearest neighbor sample matrix is ​​formed. , ,calculate Nearest Neighbor Sample Matrix Reconstruction error matrix ,in, Calculate the weight vector between the input sample and its nearest neighbors. The weight vector is then normalized to construct the input neighborhood weight matrix. , will the Input Samples and The weights of nearest neighbors are filled into the neighborhood weight matrix of the input, while the weights between non-nearest neighbor input samples are 0.

[0015] S2.3, calculate using the k-nearest neighbor method Neighborhood weight matrix .

[0016] Furthermore, step S3 specifically includes the following steps: S3.1, Calculate the matrix: ; ; in, The dimension is The identity matrix, for transpose, for The transpose of .

[0017] S3.2, for the first One latent variable, initializing the input and output weight vectors. , .

[0018] S3.3, Constructing the matrix: ; ; in, ; ; ; in, For the reason The block matrix formed, where , For the reason The block matrix formed, where , and These are the dynamic order of the output and input data, respectively.

[0019] S3.4, using weight vectors and Calculation parameters , : ; ; in, , , The dimension is The identity matrix, The dimension is The identity matrix, Represents pseudo-rebellion, For Kronecker product, for transpose, for The transpose of .

[0020] According to parameters and Update weight vector and , to obtain vector and : ; ; in, The dimension is The identity matrix, The dimension is The identity matrix; ; ; in, for transpose, for The transpose of .

[0021] Furthermore, step S3 also includes the following steps: S3.5, perform multiple iterative calculations on step S3.4 until the parameters are obtained. and convergence.

[0022] S3.6, Calculate the score vectors of the input and output. , : ; .

[0023] Calculate the load vectors of input and output , : ; ; in, yes transpose, yes The transpose of .

[0024] Scaling the data matrix: ; ; in, for transpose, for The transpose of .

[0025] S3.7, let Return to step S3.2 to extract the next set of parameters, and continue until all parameters have been extracted. Group parameters.

[0026] S3.8, will extract The input and output load matrices consist of several latent variables. , and the weight matrices of the input and output. , : ; ; ; .

[0027] Furthermore, step S4 specifically includes the following steps: S4.1 Using the load matrix and weight matrix obtained in step S3.8, calculate the projection direction matrices of the input and output. , : ; .

[0028] S4.2, Calculate the dynamic score matrix of input and output. , : ; .

[0029] Calculate the residual matrices of the input and output. , : ; .

[0030] S4.3, the dynamic score matrix of input and output is written in the following form: ; ; in, and These are the first and second inputs and outputs, respectively. A score vector, .

[0031] Calculate the first Dynamic prediction scores of input and output for each sample , : ; ; in, , This represents a diagonal matrix.

[0032] Constructing dynamic prediction score matrices for input and output , : ; .

[0033] Furthermore, step S4 also includes the following steps: S4.4, Taking the dynamic score matrix of input and output A sample, denoted as and : ; .

[0034] Calculate the dynamic score prediction error matrix of the input and output. , : ; .

[0035] S4.5, Preserve the residual matrix and After The number of samples is denoted as the residual matrix. and To ensure that the residual matrix corresponds to the sample in the dynamic score prediction error, the feature matrices of the input and output are constructed. , : ; .

[0036] S4.6 Standardize the input feature matrix by calculating the mean and standard deviation of each variable. Subtract the mean from the standard deviation of each variable and divide the result by the standard deviation. Standardize the output feature matrix by calculating the covariance matrix of the standardized input feature matrix, denoted as S4.6. and calculate rank The control limits for the reduced-rank Mahalanobis distance index are given by the chi-square distribution, denoted as . The meaning is to have On a chi-square distribution with 1 degree of freedom Quantile limit To determine the significance level, calculate the covariance matrix and rank of the standardized output feature matrix, denoted as . and The control limits for the reduced-rank Mahalanobis distance index are given by the chi-square distribution, denoted as . Meaning to have On a chi-square distribution with 1 degree of freedom Quantile limit The significance level is indicated by .

[0037] Furthermore, step S5 specifically involves: in order to detect the first... Whether a fault occurs at any time, collect the data. Time to the Data between moments, Based on whether the variable belongs to the process input variable or the process output variable, the data is divided into input data and output data, which are used as test data. The measurement variables in the test data correspond to the measurement variables in the training data in step S1.

[0038] Furthermore, step S6 specifically includes the following steps: S6.1, Standardize the test data. Using the mean and standard deviation of the input data from the training data, subtract the corresponding mean from each variable in the input data, and divide the result by the corresponding standard deviation. Similarly, using the mean and standard deviation of the output data from the training data, subtract the corresponding mean from each variable in the output data, and divide the result by the corresponding standard deviation. Denote the standardized input data as... The standardized output data is denoted as ,in, For the number of samples, It is the first Input samples at time 1 It is the first Output samples at each time point .

[0039] S6.2, using the load matrix and projection direction matrix obtained in steps S3 and S4, calculate the dynamic score matrix of the input and output of the test data. , : ; .

[0040] Calculate the residual matrix of the input and output test data. , : ; .

[0041] The residual matrices of the input and output can be written in the following form: ; ; in, and These are the first and second inputs and outputs, respectively. Each residual vector .

[0042] S6.3, the dynamic score matrix of the test data input and output is written in the following form: ; ; in, and These are the first and second inputs and outputs, respectively. A score vector, .

[0043] Calculate the first Dynamic prediction scores of input and output for each sample , : ; .

[0044] S6.4, Calculate the first... Dynamic score prediction error vector of input and output at time step , : ; .

[0045] Furthermore, step S6 also includes the following steps: S6.5, will the first The residual vector of the input at time step and dynamic score prediction error vector Combine to construct the input feature vector , will the The residual vector of the output at time step and dynamic score prediction error vector Combine to construct the output feature vector : ; ; Input feature vectors of test data Standardization is performed by using the mean and standard deviation of the feature matrix of the input data in step S4.5. For each variable, the mean is subtracted from the corresponding mean, and the result is divided by the corresponding standard deviation to output the feature vector of the test data. Standardize the vectors, and denote the standardized vectors as follows: , The reduced-rank Mahalanobis distance index of the input and the reduced-rank Mahalanobis distance index are calculated using the covariance matrix of the feature matrix in step S4.5. , : ; ; in, and These are the covariance matrices and The false rebellion.

[0046] S6.6, compare the reduced-rank Mahalanobis distance index with the control limits in step S4.5, when... and At that time, a fault occurred that affected the input but not the output. After the fault occurred, the process input was changed to suppress the fault and stabilize the output at the set value under the regulation of the closed-loop controller; when and At that time, a fault occurred that affected both the input and output. After the fault occurred, the process input changed under the regulation of the closed-loop controller, but the fault could not be completely suppressed, and the output could not stabilize at the set value; when and At that time, a fault occurred that did not affect the input but affected the output. After the fault occurred, the closed-loop controller did not make any adjustments. The process input remained unchanged, but the output could not be stabilized at the set value due to the fault.

[0047] The present invention has the following beneficial effects: The method provided by this invention uses data from normal operating conditions for modeling, eliminating the need for precise analytical models and making it widely applicable. It divides the detection space into input and output spaces, facilitating the analysis of the location of faults and determining whether they affect product quality in industrial processes. Attached Figure Description

[0048] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort. In the drawings: Figure 1 A flowchart of a closed-loop process monitoring method based on improved dynamic latent variable analysis according to the present invention is shown.

[0049] Figure 2 A schematic diagram of the input reduced-rank Mahalanobis distance index results for reactor cooling water flow velocity step change fault IDV(4) detection using the method provided by the present invention is shown.

[0050] Figure 3 A schematic diagram of the output reduced-rank Mahalanobis distance index results for reactor cooling water flow velocity step change fault IDV(4) detection using the method provided by the present invention is shown.

[0051] Figure 4A schematic diagram of the input reduced-rank Mahalanobis distance index results for condenser cooling water inlet temperature step change fault IDV(5) detection using the method provided by the present invention is shown.

[0052] Figure 5 A schematic diagram of the output reduced-rank Mahalanobis distance index results for condenser cooling water inlet temperature step change fault IDV(5) detection using the method provided by the present invention is shown.

[0053] Figure 6 A schematic diagram of the input reduced-rank Mahalanobis distance index results for reactor cooling water inlet temperature change fault IDV(11) detection using the method provided by the present invention is shown.

[0054] Figure 7 A schematic diagram of the output reduced-rank Mahalanobis distance index results for reactor cooling water inlet temperature change fault IDV(11) detection using the method provided by the present invention is shown. Detailed Implementation

[0055] The technical solution of the present invention will now be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0056] like Figure 1 The closed-loop process monitoring method based on improved dynamic latent variable analysis shown herein specifically includes the following steps: S1 collects a segment of sensor measurement data under normal operating conditions of an industrial process, and divides the sensor measurement data into input data and output data as training data; industrial processes include: chemical production processes, continuous stirred tank heaters, sewage treatment, thermal power plants, etc., which can all be considered industrial processes.

[0057] S2 standardizes the input and output data respectively, and uses the k-nearest neighbor method to calculate the neighborhood weight matrix of the input and output data respectively.

[0058] S3. Given the number of latent variables, establish an improved dynamic latent variable analysis model to obtain the input and output weight matrices and loading matrices.

[0059] S4. Calculate the projection direction matrix using the weight matrix and load matrix from step S3, construct the input and output feature matrices, standardize the feature matrices, and calculate the control limits of the covariance matrix and the reduced-rank Mahalanobis distance index.

[0060] S5. Collect sensor measurement data under real-time operating conditions of the industrial process. Divide the sensor measurement data under real-time operating conditions into input and output data as test data. The measurement variables in the test data correspond to the measurement variables in the training data in step S1.

[0061] S6. The test data is standardized. The input and output feature vectors are calculated using the projection direction matrix in step S4. The feature vectors are standardized, and the reduced-rank Mahalanobis distance index is calculated. The index is compared with the control limit in step S4 to determine whether a fault has occurred.

[0062] Specifically, step S1 is as follows: Given the dynamic order of the input data and the dynamic order of the output data Calculate the dynamic order of the industrial process system that needs to be modeled. ,collect The sample data at each time point is divided into input data and output data based on whether the variable is a process input variable or a process output variable. This is a function to find the maximum value.

[0063] Specifically, step S2 includes the following steps: S2.1, For the input data matrix, calculate the mean and standard deviation of each variable. Subtract the mean from the standard deviation of each variable, and divide the result by the standard deviation to make the mean of each variable 0 and the variance 1. This standardized matrix is ​​denoted as... Similarly, after standardizing the output data matrix, it is denoted as... ;in, For the set of real numbers, The number of samples in the training data. The number of input variables. It is the first The input samples at each time point, where , The number of variables in the output. It is the first The output samples at each time step, where .

[0064] S2.2, using the k-nearest neighbor method to calculate The neighborhood weight matrix, given the number of neighborhoods is Calculate the Euclidean distance between each input sample, and find the nearest neighbor for each input sample. Each sample is taken as a nearest neighbor of the input sample, and a nearest neighbor sample matrix is ​​formed. , ,calculate Nearest Neighbor Sample Matrix Reconstruction error matrix ,in, Calculate the weight vector between the input sample and its nearest neighbors. The weight vector is then normalized to construct the input neighborhood weight matrix. , will the Input Samples and The weights of nearest neighbors are filled into the neighborhood weight matrix of the input, while the weights between non-nearest neighbor input samples are 0.

[0065] S2.3, Similarly, following the method described in step S2.2, the k-nearest neighbor method is used to calculate... Neighborhood weight matrix .

[0066] Specifically, step S3 includes the following steps: S3.1, Calculate the matrix: ; ; in, The dimension is The identity matrix, for transpose, for The transpose of .

[0067] S3.2, for the first One latent variable, initializing the input and output weight vectors. , .

[0068] S3.3, Constructing the matrix: ; ; in, ; ; ; in, For the reason The block matrix formed, where , For the reason The block matrix formed, where , and These are the dynamic order of the output and input data, respectively.

[0069] S3.4, using weight vectors and Calculation parameters , : ; ; in, , , The dimension is The identity matrix, The dimension is The identity matrix, Represents pseudo-rebellion, For Kronecker product, for transpose, for The transpose of .

[0070] According to parameters and Update weight vector and , to obtain vector and : ; ; in, The dimension is The identity matrix, The dimension is The identity matrix; ; ; in, for transpose, for The transpose of .

[0071] Specifically, step S3 also includes the following steps: S3.5, perform multiple iterative calculations on step S3.4 until the parameters are obtained. and convergence.

[0072] S3.6, Calculate the score vectors of the input and output. , : ; .

[0073] Calculate the load vectors of input and output , : ; ; in, yes transpose, yes The transpose of .

[0074] Scaling the data matrix: ; ; in, for transpose, for The transpose of .

[0075] S3.7, let Return to step S3.2 to extract the next set of parameters, and continue until all parameters have been extracted. Group parameters.

[0076] S3.8, will extract The input and output load matrices consist of several latent variables. , and the weight matrices of the input and output. , : ; ; ; .

[0077] Specifically, step S4 includes the following steps: S4.1 Using the load matrix and weight matrix obtained in step S3.8, calculate the projection direction matrices of the input and output. , : ; .

[0078] S4.2, Calculate the dynamic score matrix of input and output. , : ; .

[0079] Calculate the residual matrices of the input and output. , : ; .

[0080] S4.3, the dynamic score matrix of input and output is written in the following form: ; ; in, and These are the first and second inputs and outputs, respectively. A score vector, .

[0081] Calculate the first Dynamic prediction scores of input and output for each sample , : ; ; in, , This represents a diagonal matrix.

[0082] Constructing dynamic prediction score matrices for input and output , : ; .

[0083] Specifically, step S4 also includes the following steps: S4.4, Taking the dynamic score matrix of input and output A sample, denoted as and : ; .

[0084] Calculate the dynamic score prediction error matrix of the input and output. , : ; .

[0085] S4.5, Preserve the residual matrix and After The number of samples is denoted as the residual matrix. and To ensure that the residual matrix corresponds to the sample in the dynamic score prediction error, the feature matrices of the input and output are constructed. , : ; .

[0086] S4.6 Standardize the input feature matrix by calculating the mean and standard deviation of each variable. Subtract the mean from the standard deviation of each variable and divide the result by the standard deviation. Similarly, standardize the output feature matrix and calculate the covariance matrix of the standardized input feature matrix, denoted as S4.6. and calculate rank The control limits for the reduced-rank Mahalanobis distance index are given by the chi-square distribution, denoted as . Meaning to have On a chi-square distribution with 1 degree of freedom Quantile limit To determine the significance level, similarly calculate the covariance matrix and rank of the standardized output feature matrix, denoted as . and The control limits for the reduced-rank Mahalanobis distance index are given by the chi-square distribution, denoted as . Meaning to have On a chi-square distribution with 1 degree of freedom Quantile limit The significance level is indicated by .

[0087] Specifically, step S5 is as follows: In order to detect the first Whether a fault occurs at any time, collect the data. Time to the Data between moments, Based on whether the variable belongs to the process input variable or the process output variable, the data is divided into input data and output data, which are used as test data. The measurement variables in the test data correspond to the measurement variables in the training data in step S1.

[0088] Specifically, step S6 includes the following steps: S6.1, Standardize the test data. Using the mean and standard deviation of the input data from the training data, subtract the corresponding mean from each variable in the input data, and divide the result by the corresponding standard deviation. Similarly, using the mean and standard deviation of the output data from the training data, subtract the corresponding mean from each variable in the output data, and divide the result by the corresponding standard deviation. Denote the standardized input data as... The standardized output data is denoted as ,in, For the number of samples, It is the first Input samples at time 1 It is the first Output samples at each time point .

[0089] S6.2, using the load matrix and projection direction matrix obtained in steps S3 and S4, calculate the dynamic score matrix of the input and output of the test data. , : ; .

[0090] Calculate the residual matrix of the input and output test data. , : ; .

[0091] The residual matrices of the input and output can be written in the following form: ; ; in, and These are the first and second inputs and outputs, respectively. Each residual vector .

[0092] S6.3, the dynamic score matrix of the test data input and output is written in the following form: ; ; in, and These are the first and second inputs and outputs, respectively. A score vector, .

[0093] Calculate the first Dynamic prediction scores of input and output for each sample , : ; .

[0094] S6.4, Calculate the first... Dynamic score prediction error vector of input and output at time step , : ; .

[0095] Specifically, step S6 also includes the following steps: S6.5, will the first The residual vector of the input at time step and dynamic score prediction error vector Combine to construct the input feature vector Similarly, the first The residual vector of the output at time step and dynamic score prediction error vector Combine to construct the output feature vector : ; .

[0096] Input feature vectors of test data Standardization is performed by using the mean and standard deviation of the feature matrix of the input data from step S4.5. For each variable, subtract the corresponding mean and divide the result by the corresponding standard deviation. Similarly, the output feature vector of the test data is standardized. Standardize the vectors, and denote the standardized vectors as follows: , The reduced-rank Mahalanobis distance index of the input and the reduced-rank Mahalanobis distance index are calculated using the covariance matrix of the feature matrix in step S4.5. , : ; ; in, and These are the covariance matrices and The false rebellion.

[0097] S6.6, compare the reduced-rank Mahalanobis distance index with the control limits in step S4.5, when... and At that time, a fault occurred that affected the input but not the output. After the fault occurred, the process input was changed to suppress the fault and stabilize the output at the set value under the regulation of the closed-loop controller; when and At that time, a fault occurred that affected both the input and output. After the fault occurred, the process input changed under the regulation of the closed-loop controller, but the fault could not be completely suppressed, and the output could not stabilize at the set value; when and At that time, a fault occurred that did not affect the input but affected the output. After the fault occurred, the closed-loop controller did not make any adjustments. The process input remained unchanged, but the output could not be stabilized at the set value due to the fault.

[0098] This embodiment uses Matlab tools and the fault IDV(4), IDV(5), and IDV(11) in the Tennessee Eastman process to illustrate the present invention, and the effects of the present invention are shown in conjunction with the accompanying drawings.

[0099] The Tennessee-Eastman Process (TEP) comprises five main units: reactor, condenser, compressor, separator, and stripper. The TEP includes 12 manipulated variables and 41 measured variables (22 continuous variables and 19 component measurements). The TEP simulation includes 21 preset faults, 16 known and 5 unknown faults. The experiment used 960 data samples, with faults occurring starting from the 161st sample. This experiment selected 22 continuous measured variables and 11 manipulated variables for simulation verification. The specific process is as follows: (1) Let the manipulated variable under normal operating conditions be denoted as the input, and the continuously measured variable as the output. The dynamic order of the process is given as... The number of latent variables is ; (2) Calculate the mean and standard deviation of each input variable. Subtract the mean from the input variable and divide the result by the standard deviation. Similarly, calculate the mean and standard deviation of each output variable. Subtract the mean from the output variable and divide the result by the standard deviation to obtain the standardized input matrix and output matrix. , ,in , , Each row represents a sample, and each column represents a variable. It is the first Input samples at time 1 It is the first Output samples at each time point .

[0100] The k-nearest neighbor method is used to calculate the neighborhood weight matrices of the standardized input and output, respectively. , .

[0101] (3) For a given number of latent variables The weight matrices of input and output are calculated using an improved dynamic latent variable analysis method. , and load matrix , The specific steps are as follows: For the first latent variable, initialize its weight vector, calculate the parameters of the dynamic model, update the weight vector according to the parameters, and repeat the iteration to calculate the model parameters and update the weight vector until the weight vector converges. Calculate the load vector and scale the input and output data. Then, extract the weight vectors and model parameters of the second, third, and fourth latent variables in sequence, and construct the required matrix using the weight vectors and load vectors.

[0102] (4) Calculate the input projection direction matrix using the input weight matrix and load matrix. Similarly, calculate the output projection direction matrix. Calculate the residual matrix and dynamic score prediction error matrix of the input and output, respectively, and combine them into the feature matrices of the input and output. Standardize the feature matrices and calculate the covariance matrix of the feature matrices of the input and output respectively. , And calculate the corresponding rank. , Give the control limits for the reduced-rank Mahalanobis distance index of the input and output. , , The significance level is indicated by .

[0103] (5) The test data is divided into input and output data according to the variable grouping method of the training data. The input of the test data is standardized using the mean and standard deviation of the input of the training data. Similarly, the output of the test data is standardized using the mean and standard deviation of the output of the training data.

[0104] (6) Using the input and output weight matrices, load matrices, and projection direction matrices obtained during the offline modeling process, calculate the residual matrix and dynamic score prediction error matrix of the test data, and combine them into the input feature matrix and the output feature matrix. Using the mean and variance of the feature matrices obtained during the offline process, standardize the input and output feature matrices respectively, and calculate the reduced-rank Mahalanobis distance index of the input and output: ; ; in, and These are the covariance matrices and The false rebellion, For standardization The input feature vector at time step 1. For standardization The output feature vector at time t, .

[0105] (7) Compare the reduced-rank Mahalanobis distance indices of the input and output with their corresponding control limits to determine whether a fault has occurred. Figure 2-7 The results of this invention on the detection of fault IDV(4), IDV(5), and IDV(11) are presented.

[0106] When an IDV(4) fault occurs, the temperature in the reactor will suddenly rise, which needs to be compensated by closed-loop control. After the fault occurs, other measured and control variables remain stable. The fault will affect the output variable at the beginning, but under the regulation of the closed-loop controller, the output will basically stabilize at the set value, so the output space will alarm when the fault occurs. Due to the regulation of the controller, the process input variable changes, so the input space will continue to alarm, and the fault is judged to be a fault that does not affect the output. The fault detection rate (FDR) of IDV(4) of this invention is 100.00%, and the false alarm rate (FAR) is 1.27%.

[0107] When IDV(5) fails, the flow rate at the condenser outlet increases, leading to an increase in the separator temperature and a change in the separator's cooling water outlet temperature. This failure is compensated for by the control loop adjustment, as are the trends of other variables. IDV(5) causes continuous abnormalities in the process input, but the measured variables slowly return to their normal operating range, thus triggering a continuous alarm in the input space, while the output space stabilizes and no longer triggers an alarm. The present invention achieves a detection rate of 97.25% for IDV(5) and a false negative rate of 1.27%.

[0108] IDV(11) causes random variations in the reactor cooling water inlet temperature. This fault can cause fluctuations in the reactor cooling water flow rate and affect the reactor temperature. Other variables remain relatively stable under normal operating conditions. Because IDV(11) is a random fault, it is difficult for the closed-loop controller to regulate it, so both process input and output variables are affected by the fault, and both input and output spaces will trigger alarms after the fault occurs. The detection rate of IDV(11) in this invention is 66.13%, and the false negative rate is 0.00%.

[0109] Compared to traditional principal component analysis, canonical correlation analysis, and dynamic internal principal component analysis based on dynamic latent variable analysis, this invention provides better detection results for closed-loop processes and can determine whether the type of fault will affect the output.

[0110] Of course, the above description is not intended to limit the present invention, and the present invention is not limited to the examples given above. Any changes, modifications, additions or substitutions made by those skilled in the art within the scope of the present invention should also fall within the protection scope of the present invention.

Claims

1. A closed-loop process monitoring method based on improved dynamic latent variable analysis, characterized in that, Specifically, the steps include the following: S1: Collect a segment of sensor measurement data under normal operating conditions of the industrial process, and divide the sensor measurement data into input data and output data as training data; S2, standardize the input and output data respectively, and use the k-nearest neighbor method to calculate the neighborhood weight matrix of the input and output data respectively; S3, Given the number of latent variables, establish an improved dynamic latent variable analysis model to obtain the input and output weight matrix and loading matrix; S4. Calculate the projection direction matrix using the weight matrix and load matrix from step S3, construct the input and output feature matrices, standardize the feature matrices, and calculate the control limits of the covariance matrix and the reduced-rank Mahalanobis distance index. S5. Collect sensor measurement data under real-time operating conditions of the industrial process, divide the sensor measurement data under real-time operating conditions into input and output data, and use them as test data. The measurement variables in the test data correspond to the measurement variables in the training data in step S1. S6. The test data is standardized. The input and output feature vectors are calculated using the projection direction matrix in step S4. The feature vectors are standardized, and the reduced-rank Mahalanobis distance index is calculated. The index is compared with the control limit in step S4 to determine whether a fault has occurred. Step S1 is as follows: The dynamic order of the given input data and the dynamic order of the output data Calculate the dynamic order of the industrial process system that needs to be modeled. ,collect The sample data at each time point is divided into input data and output data based on whether the variable is a process input variable or a process output variable. This is a function to find the maximum value. Step S2 specifically includes the following steps: S2.1, For the input data matrix, calculate the mean and standard deviation of each variable. Subtract the mean from the standard deviation of each variable, and divide the result by the standard deviation to make the mean of each variable 0 and the variance 1. This standardized matrix is ​​denoted as... After standardizing the output data matrix, it is denoted as... ;in, For the set of real numbers, The number of samples in the training data. The number of input variables. It is the first The input samples at time t, where , The number of variables in the output. It is the first The output samples at each time step, where ; S2.2, using the k-nearest neighbor method to calculate The neighborhood weight matrix, given the number of neighborhoods is Calculate the Euclidean distance between each input sample, and find the nearest neighbor for each input sample. Each sample is taken as a nearest neighbor of the input sample, and a nearest neighbor sample matrix is ​​formed. , ,calculate Nearest Neighbor Sample Matrix Reconstruction error matrix ,in, Calculate the weight vector between the input sample and its nearest neighbors. The weight vector is then normalized to construct the input neighborhood weight matrix. , will the Input Samples and The weights of nearest neighbors are filled into the neighborhood weight matrix of the input, and the weights between non-near neighbors are 0. S2.3, calculate using the k-nearest neighbor method Neighborhood weight matrix ; Step S3 specifically includes the following steps: S3.1, Calculate the matrix: ； ； in, The dimension is The identity matrix, for transpose, for Transpose of; S3.2, for the first One latent variable, initializing the input and output weight vectors. , ; S3.3, Constructing the matrix: ； ； in, ; ； ； in, For the reason The block matrix formed, where , For the reason The block matrix formed, where , and These are the dynamic order of the output and input data, respectively; S3.4, using weight vectors and Calculation parameters , : ； ； in, , , The dimension is The identity matrix, The dimension is The identity matrix, Represents false rebellion, For Kronecker product, for transpose, for Transpose of; According to parameters and Update weight vector and , to obtain vector and : ； ； in, The dimension is The identity matrix, The dimension is The identity matrix; ； ； in, for transpose, for Transpose of; Step S3 also includes the following steps: S3.5, Perform multiple iterative calculations on step S3.4 until the parameters are obtained. and convergence; S3.6, Calculate the score vectors of the input and output. , : ； ； Calculate the load vectors of input and output , : ； ； in, yes transpose, yes Transpose of; Scaling the data matrix: ； ； in, for transpose, for Transpose of; S3.7, let Return to step S3.2 to extract the next set of parameters, and continue until all parameters have been extracted. Group parameters; S3.8, extract The input and output load matrices consist of several latent variables. , and the weight matrices of the input and output. , : ； ； ； ； Step S4 specifically includes the following steps: S4.1 Using the load matrix and weight matrix obtained in step S3.8, calculate the projection direction matrices of the input and output. , : ； ； S4.2, Calculate the dynamic score matrix of input and output. , : ； ； Calculate the residual matrices of the input and output. , : ； ； S4.3, the dynamic score matrix of input and output is written in the following form: ； ； in, and These are the first and second inputs and outputs, respectively. A score vector, ; Calculate the first Dynamic prediction scores of input and output for each sample , : ； ； in, , Represents a diagonal matrix; Construct dynamic prediction score matrices for input and output. , : ； ； Step S4 also includes the following steps: S4.4, Taking the dynamic score matrix of input and output A sample, denoted as and : ； ； Calculate the dynamic score prediction error matrix of the input and output. , : ； ； S4.5, Preserve the residual matrix and After The number of samples is denoted as the residual matrix. and To ensure that the residual matrix corresponds to the sample in the dynamic score prediction error, the feature matrices of the input and output are constructed. , : ； ； S4.6 Standardize the input feature matrix by calculating the mean and standard deviation of each variable. Subtract the mean from the standard deviation of each variable and divide the result by the standard deviation. Standardize the output feature matrix by calculating the covariance matrix of the standardized input feature matrix, denoted as S4.

6. and calculate rank The control limits for the reduced-rank Mahalanobis distance index are given by the chi-square distribution, denoted as . The meaning is to have On a chi-square distribution with 1 degree of freedom Quantile limit To determine the significance level, calculate the covariance matrix and rank of the standardized output feature matrix, denoted as . and The control limits for the reduced-rank Mahalanobis distance index are given by the chi-square distribution, denoted as . The meaning is to have On a chi-square distribution with 1 degree of freedom Quantile limit The significance level; Step S5 specifically involves: In order to detect the first Whether a fault occurs at any time, collect the data. Time to the Data between moments, Based on whether the variable belongs to the process input variable or the process output variable, the data is divided into input data and output data, which are used as test data. The measurement variables in the test data correspond to the measurement variables in the training data in step S1. Step S6 specifically includes the following steps: S6.1, Standardize the test data. Using the mean and standard deviation of the input data from the training data, subtract the corresponding mean from each variable in the input data, and divide the result by the corresponding standard deviation. Similarly, using the mean and standard deviation of the output data from the training data, subtract the corresponding mean from each variable in the output data, and divide the result by the corresponding standard deviation. Denote the standardized input data as... The standardized output data is denoted as ,in, For the number of samples, It is the first Input samples at time 1 It is the first Output samples at each time point ; S6.2, using the load matrix and projection direction matrix obtained in steps S3 and S4, calculate the dynamic score matrix of the input and output of the test data. , : ； ； Calculate the residual matrix of the input and output test data. , : ； ； The residual matrices of the input and output can be written in the following form: ； ； in, and These are the first and second inputs and outputs, respectively. Each residual vector ; S6.3, the dynamic score matrix of the test data input and output is written in the following form: ； ； in, and These are the first and second inputs and outputs, respectively. A score vector, ; Calculate the first Dynamic prediction scores of input and output for each sample , : ； ； S6.4, Calculate the first... Dynamic score prediction error vector of input and output at time step , : ； ； Step S6 also includes the following steps: S6.5, will the first The residual vector of the input at time step and dynamic score prediction error vector Combine to construct the input feature vector , will the The residual vector of the output at time step and dynamic score prediction error vector Combine to construct the output feature vector : ； ； Input feature vectors of test data Standardization is performed by using the mean and standard deviation of the feature matrix of the input data in step S4.

5. For each variable, the mean is subtracted from the corresponding mean, and the result is divided by the corresponding standard deviation to output the feature vector of the test data. Standardize the vectors, and denote the standardized vectors as follows: , The reduced-rank Mahalanobis distance index of the input and the reduced-rank Mahalanobis distance index are calculated using the covariance matrix of the feature matrix in step S4.

5. , : ； ； in, and These are the covariance matrices. and The false reversal; S6.6, compare the reduced-rank Mahalanobis distance index with the control limits in step S4.5, when... and At that time, a fault occurred that affected the input but not the output. After the fault occurred, the process input was changed to suppress the fault and stabilize the output at the set value under the regulation of the closed-loop controller; when and At that time, a fault occurred that affected both the input and output. After the fault occurred, the process input changed under the regulation of the closed-loop controller, but the fault could not be completely suppressed, and the output could not stabilize at the set value; when and At that time, a fault occurred that did not affect the input but affected the output. After the fault occurred, the closed-loop controller did not make any adjustments. The process input remained unchanged, but the output could not be stabilized at the set value due to the fault.

Images