A method for detecting abnormal operation of a converter based on time sequence characteristic analysis

Abstract

The application discloses a kind of based on timing characteristic analysis one section converter operating abnormality detection method, to be in by analyzing the timing characteristic of one section converter sampling data, complete the real-time abnormality detection of one section converter operating state.Specifically, the method of the present application carries out timing characteristic analysis on multiple time nodes for the sample data collected when one section converter is in normal operation, establishes data model for online real-time abnormality detection after learning corresponding timing characteristic, and then detects whether the operation of one section converter is abnormal through a comprehensive abnormality detection index.Compared with prior art, the method of the present application obtains corresponding feature transformation matrix through timing characteristic analysis process capable of minimizing the difference between score vectors in time sequence, and completes the real-time online abnormality detection of one section converter on the basis of monitoring through the comprehensive abnormality detection index of three different abnormality detection indexes.

Description

TECHNICAL FIELD

[0001] The present application relates to a chemical equipment operation state monitoring method, in particular to a one-stage reformer operation abnormality detection method based on time sequence feature analysis. BACKGROUND

[0002] In the production process of synthetic ammonia, hydrogen gas, as one of the important raw materials, is converted from natural gas, and the natural gas conversion unit includes three parts: pre-reformer, one-stage reformer and two-stage reformer. However, the conversion reaction mainly occurs in the one-stage reformer, so the monitoring and control of the operation state of the device is particularly important. The stable operation state can effectively ensure the purity of hydrogen gas, and thus ensure the yield of synthetic ammonia. With the wide application of computer technology and advanced instruments, the one-stage reformer is equipped with multiple measuring instruments that can measure sample data such as temperature, pressure and flow rate in real time at fixed time intervals, which provides a solid data basis for implementing data-driven one-stage reformer operation abnormality detection.

[0003] A Chinese patent application for a synthetic ammonia one-stage reformer operation abnormality detection method based on time sequence difference feature extraction, with application number 202210651302X and title, discloses a synthetic ammonia one-stage reformer operation abnormality detection method. The method extracts the time sequence difference features in the latest sample data at the latest sampling time, and detects the abnormality in the one-stage reformer operation through the abnormal change of the difference features. Since the method regards all the latest sample data as potential abnormal data, it needs real-time feature value problem solving, which not only increases the computational load of the corresponding abnormality detection device implementation method, but also is easily affected by the measurement noise in the sample data, resulting in more unnecessary false alarms of abnormal states, causing unnecessary trouble for the site operators.

[0004] Considering that the stable operation of the one-stage reformer will inevitably lead to a greater correlation between the sample data reflected in the time sequence, when an abnormal operation state occurs, the time sequence features reflected in the sampling time should also change abnormally. In order to avoid excessive false alarms of abnormal states, the time sequence features of the sample data representing the normal or expected operation state of the one-stage reformer can be analyzed to complete the abnormality detection of the one-stage reformer operation state. SUMMARY

[0005] Compared to existing technologies, the main technical problem this invention aims to solve is: how to analyze the temporal characteristics of sampled data from a converter segment to achieve real-time anomaly detection of the converter's operating status. Specifically, the method of this invention performs temporal characteristic analysis on sample data collected from a converter segment during its normal operation at multiple time points. After learning to represent the corresponding temporal characteristics, a data model for online real-time anomaly detection is established. Finally, a comprehensive anomaly detection index is used to detect whether anomalies have occurred in the operation of the converter segment.

[0006] The technical solution adopted by the present invention to solve the above problems is: a method for detecting abnormal operation of a single-stage converter based on time-series feature analysis, comprising the following steps.

[0007] Step (1): After determining the measurement variables of a section of the converter, while the section of the converter is operating normally, sample data at least 800 sampling times are continuously collected for the measurement variables to form a data matrix X. The determined measurement variables are as follows: natural gas feed flow rate (F01), fuel exhaust gas feed flow rate (F02), heat exchanger outlet pressure (P01), furnace gas pressure (P02), heat exchanger outlet temperature (T01), preheated natural gas temperature (T02), mixed gas feed temperature (T03), furnace left top temperature (T04), furnace right top temperature (T05), furnace top mixed gas temperature (T06), furnace left bottom outlet temperature (T07), furnace right bottom outlet temperature (T08), and converted gas outlet temperature (T09). Each row vector in the data matrix X is composed of the sample data of the measurement variables at the same sampling time arranged in a fixed measurement variable order. The first row vector to the last row vector in X are arranged in the order of sampling time.

[0008] Step (2): Calculate the average values ​​μ1, μ2, ..., μ of the first, second, to thirteenth column vectors in the data matrix X. 13 and standard deviation δ1,δ2,…,δ 13 Then, the mean and standard deviation are used to standardize each row vector in X, and the standardized row vectors are then combined into a matrix.

[0009] Step (3): After setting the time series order D, the matrix The (D+1)th row vector to the last row vector forms a 0th-order time sequence matrix. matrix The first-order time sequence matrix is ​​formed by the D-th row vector to the second-to-last row vector. matrix The second-order time sequence matrix is ​​formed by the (D-1)th row vector to the third-to-last row vector. According to this rule until the D order time series matrix is obtained

[0010] Step (4): For the 0 order time series matrix 1 order time series matrix 2 order time series matrix to D order time series matrix Implementing time series feature analysis, obtaining feature transformation matrix P.

[0011] Step (5): Set the time series feature parameter K equal to a positive integer less than 13, and then form the dynamic transformation matrix P1 by the first column vector to the Kth column vector in the feature transformation matrix P, form the static transformation matrix P2 by the K+1th column vector to the 13th column vector in the feature transformation matrix P, and then calculate the 0 order score matrix S0, the 1 order score matrix S1, the 2 order score matrix S2, …, the D order score matrix S and respectively, to obtain the 0 order score matrix S0, the 1 order score matrix S1, the 2 order score matrix S2, …, the D order score matrix S D and the static score matrix U; wherein d = 0, 1, 2, …, D.

[0012] Step (6): Determine the regression model between the 0 order score matrix S0 and S1, S2, …, S D using the least squares regression algorithm: wherein the input matrix the regression coefficient matrix and the error matrix The superscript T represents the transpose of the matrix or vector.

[0013] Step (7): Calculate the error anomaly detection index vector Q, the dynamic anomaly detection index vector G and the static anomaly detection index vector H according to the formulas Q = diag{EE T}, G = diag{SS T} and H = diag{UU T} respectively, and then set the error anomaly control limit the dynamic anomaly control limit and the static anomaly control limit respectively corresponding to the maximum data in Q, G and H; wherein diag{} represents the operation of converting the diagonal elements in the matrix within the curly braces into a column vector, and the superscript T represents the transpose of the matrix or vector.

[0014] Step (8): Calculate the comprehensive anomaly detection index vector Φ according to the formula and then determine the corresponding upper control limit according to the data in Φ using the three-sigma rule.

[0015] Step (9): After the parameters for online real-time anomaly detection are reserved, the following steps (9.1) to (9.4) are repeatedly executed to implement online anomaly detection for a section of a reformer; wherein the reserved parameters include: the average values μ1, μ2, …, μ 13 and the standard deviations δ1, δ2, …, δ 13 , the dynamic transformation matrix P1 and the static transformation matrix P2 in step (5), the regression coefficient matrix B in step (6), the error anomaly control limit dynamic anomaly control limit and static anomaly control limit , and the control upper limit

[0016] Step (9.1) collects sample data of 13 measurement variables at the latest sampling time t of a section of a reformer, and arranges the sample data in the order of the fixed measurement variables in step (1) to form a 1 × 13 dimensional data vector x t After that, the average values μ1, μ2, …, μ 13 and the standard deviations δ1, δ2, …, δ 13 are used to perform standardization processing on the data vector x t , so as to correspondingly obtain a row vector

[0017] Step (9.2) calculates the 0-order score vector s0, the 1-order score vector s1, the 2-order score vector s2 to the D-order score vector s D respectively according to the formula After that, the static score vector u t at the latest sampling time t is calculated according to the formula ; wherein represents the row vector corresponding to the data vector x t-d collected and composed at the d sampling times before t after standardization processing, and when d = 0, represents

[0018] Step (9.3) combines s1, s2, …, s D into an input vector After that, the error vector e t is calculated according to the formula

[0019] Step (9.4) calculates the comprehensive anomaly detection index at the latest sampling time t according to the formula After that, it is judged whether is greater than ​If yes, the first-stage reformer operation is abnormal; if no, the first-stage reformer operation is normal.

[0020] In the above steps of the present application, step (4) is the key to implement the present application, and its specific implementation process includes the following steps (4.1) to (4.4).

[0021] Step (4.1): obtaining a matrix performing singular value decomposition to obtain C=VΛV T and then calculating an orthogonal transformation matrix Θ according to the formula Θ=VΛ -1 / 2 ; wherein V represents a unitary matrix generated by singular value decomposition, and Λ is a diagonal matrix composed of non-zero singular values.

[0022] Step (4.2): calculating a 0-order characteristic matrix Z0, a 1-order characteristic matrix Z1, a 2-order characteristic matrix Z2, and a D-order characteristic matrix Z D .

[0023] Step (4.3): calculating a characteristic correlation matrix M according to the formula , and then performing singular value decomposition on M to obtain M=NAN T ; wherein A is a diagonal matrix composed of non-zero singular values arranged in descending order of numerical value, and N is a unitary matrix generated by singular value decomposition.

[0024] Step (4.4): calculating a characteristic transformation matrix P according to the formula P=ΘN.

[0025] The purpose of the sequence feature analysis in the implementation of the present application is to transform the 0-order to D-order sequence matrices into corresponding score vectors s1, s2, …, s D through the transformation of the characteristic transformation vectors p, and then the score vectors change slowly, i.e., the difference between the score vectors in time sequence is minimized, as shown in formula (1).

[0026]

[0027] wherein the objective function in formula (1) The constraint condition of the minimization problem in formula (1) is that the corresponding score vectors are obtained by the transformation of different characteristic transformation vectors p, and their lengths are equal to 1 and mutually orthogonal. Therefore, minimizing is equivalent to maximizing

[0028] In summary, the following maximization problem can be defined to solve multiple characteristic transformation vectors:

[0029] ​

[0030] Where st is an abbreviation for Subject To, representing the constraint condition, and I represents the identity matrix. Solving equation (2) can be transformed into a classic generalized eigenvalue problem: Where φ represents the diagonal matrix composed of eigenvalues.

[0031] Solving this generalized eigenvalue problem can be divided into two steps. The first step can be... The first step is to perform singular value decomposition, i.e., the process of step (4.1); the second step is to... The eigencorrelation matrix between the transformed eigenma matrices of order 0 to order D Perform singular value decomposition, which is the process of step (4.3).

[0032] Therefore, after performing time series feature analysis through step (4) of the present invention, the corresponding feature transformation matrix P can be obtained. Since the non-zero singular values ​​on the diagonal of A are arranged in descending order of numerical value, the feature transformation vectors of the first few columns of P can be transformed to obtain dynamic features with little difference in time sequence, while the feature transformation vectors of the last few columns of P cannot reflect the slow change in time sequence, and are therefore called static features.

[0033] As a further improvement of the present invention, the setting of the timing order D in step (3) has a direct impact on the temporal variation characteristics of the sample data of a converter segment. Generally, the timing order D can be set according to the sampling frequency of a converter segment, and its preferred setting range is 2≤D≤8. The larger the sampling frequency, the larger the timing order D. Of course, the best way to set the timing order D is as shown in steps (3.1) to (3.4):

[0034] Step (3.1): Set α = 0, D = 2 and the threshold ε to a positive number not greater than 0.1.

[0035] Step (3.2): Perform the steps in step (3) to form the 0th order time sequence matrix. 1st order time sequence matrix 2nd order time matrix up to D-order time sequence matrix The implementation process.

[0036] Step (3.3): After obtaining the diagonal matrix A by executing steps (4.1) to (4.2), set β to be equal to the sum of the diagonal elements in A.

[0037] Step (3.4): Determine whether β-α is greater than ε; if yes, set D = D+1 and return to step (3.2); if no, obtain the timing order D required in step (3).

[0038] Through steps (3.1) to (3.4), the optimal time series order D can be objectively determined, avoiding the reduction of anomaly detection performance when implementing online anomaly detection due to human subjective experience errors in setting the time series order D.

[0039] As a further improvement to step (5) in this invention, when setting the timing feature parameter K, it is preferable to determine the timing feature parameter K to be set according to steps (5.1) to (5.2) as shown below.

[0040] Step (5.1): Denote the elements on the diagonal of the diagonal matrix A as a1, a2, ..., a F Then, set f = 1; where the subscript F is equal to the number of row vectors in the diagonal matrix A.

[0041] Step (5.2): According to η f = (a1+a2+…+a) f ) / (a1+a2+…+a F Calculate the cumulative probability η f Then, determine η f Is it less than 0.85? If yes, set f = f + 1 and repeat step (5.2); if no, set the timing characteristic parameter K = f and then complete the setting of the timing characteristic parameter K.

[0042] Through steps (5.1) to (5.2), the corresponding time-series characteristic parameters K can be objectively determined without being affected by the interference of parameters set by human subjective experience. This avoids reducing the anomaly detection performance when implementing online anomaly detection due to incorrect setting of inappropriate time-series characteristic parameters K. Attached Figure Description

[0043] Figure 1 This is a flowchart illustrating the implementation of the method of the present invention.

[0044] Figure 2 This is a schematic diagram of the process flow of a single-stage converter. Detailed Implementation

[0045] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments.

[0046] like Figure 1 As shown in the figure, the present invention discloses a method for detecting abnormal operation of a single-stage converter based on time-series feature analysis. The specific implementation of the method of the present invention will be described below with reference to specific application examples.

[0047] After determining the measurement variables of a converter section in step (1), when the converter section is operating normally, sample data of 800 sampling times are continuously collected for the measurement variables to form a data matrix X. Each row vector in the data matrix X is composed of the sample data of the determined measurement variables at the same sampling time arranged in a fixed order of measurement variables. The first row vector to the last row vector in X is arranged in the order of sampling times.

[0048] In such Figure 2 In the section of the converter shown, the determined measurement variables are arranged in a fixed order as follows: natural gas feed flow rate F01, fuel exhaust gas feed flow rate F02, heat exchanger outlet pressure P01, furnace gas pressure P02, heat exchanger outlet temperature T01, preheated natural gas temperature T02, mixed gas feed temperature T03, furnace left top temperature T04, furnace right top temperature T05, furnace top mixed gas temperature T06, furnace left bottom outlet temperature T07, furnace right bottom outlet temperature T08, and converted gas outlet temperature T09.

[0049] It should be noted that the aforementioned fixed order of measurement variables is used to explain the fixed order of measurement variables used in this invention, rather than restricting the fixed order of measurement variables. In fact, different orders of measurement variables will not affect the subsequent anomaly detection performance. Therefore, the order of measurement variables is not unique; it is only necessary to ensure that the order of measurement variables used in succession is consistent.

[0050] Step (2) Calculate the average values ​​μ1, μ2, ..., μ of the first, second, to thirteenth column vectors in the data matrix X. 13 and standard deviation δ1,δ2,…,δ 13 Then, the mean and standard deviation are used to standardize each row vector in X, and the standardized row vectors are then combined into a matrix.

[0051] In step (2) above, μ1 and δ1 are the mean and standard deviation of all data in the first column vector of X, μ2 and δ2 are the mean and standard deviation of all data in the second column vector of X, and μ 13 and δ 13 Let x be the mean and standard deviation of all data in the 13th column vector of X. The standardization process is common knowledge in the field, and its specific implementation is as follows: Let x be equal to the row vector of any row in X, and then use the formula... Standardizing x yields the standardized row vector. Where the average value vector μ = [μ1, μ2, ..., μ 13 The elements on the diagonal of the diagonal matrix θ are δ1, δ2, ..., δ13 Composition.

[0052] After setting the order of time series D = 3 in step (3), the matrix is composed of the 4th row vector to the 800th row vector in the matrix is composed of the 3rd row vector to the 799th row vector in the matrix is composed of the 2nd row vector to the 798th row vector in the matrix is composed of the 1st row vector to the 797th row vector in the matrix is composed of the 0th row vector to the 796th row vector in the matrix is composed of the -1st row vector to the 795th row vector in the matrix is composed of the -2nd row vector to the 794th row vector in the matrix

[0053] In step (4), the 0th order time series matrix the 1st order time series matrix the 2nd order time series matrix and the 3rd order time series matrix are subjected to time series feature analysis to obtain a feature transformation matrix P, which includes the following steps (4.1) to (4.4).

[0054] Step (4.1): Singular value decomposition is performed on the matrix to obtain C = VΛV T , and then the orthogonal transformation matrix Θ is calculated according to the formula Θ = VΛ -1 / 2 ; wherein V represents the unitary matrix generated by singular value decomposition, and Λ is a diagonal matrix composed of non-zero singular values.

[0055] Step (4.2): The 0th order feature matrix Z0, the 1st order feature matrix Z1, the 2nd order feature matrix Z2, and the Dth order feature matrix Z are calculated respectively according to the formula D .

[0056] Step (4.3): The feature correlation matrix M is calculated according to the formula , and then singular value decomposition is performed on M to obtain M = NAN T ; wherein A is a diagonal matrix composed of non-zero singular values arranged in descending order of numerical value, and N is a unitary matrix generated by singular value decomposition.

[0057] Step (4.4): The feature transformation matrix P is calculated according to the formula P = ΘN.

[0058] In step (5), the time series feature parameter K = 8 is set, and then the 1st column vector to the K = 8th column vector in the feature transformation matrix P are composed into a dynamic transformation matrix P1, and the 9th column vector to the 13th column vector in P are composed into a static transformation matrix P2, and then the dynamic transformation matrix P1 and the static transformation matrix P2 are subjected to feature analysis to obtain a feature transformation matrix P according to the formula and The 0-order score matrix S0, the 1-order score matrix S1, the 2-order score matrix S2, the 3-order score matrix S3 and the static score matrix U are calculated respectively.

[0059] Step (6) determines the regression model between S0 and S1, S2, S3 by using the least square regression algorithm: Wherein, the input matrix is The regression coefficient matrix is And the error matrix is The superscript T represents the transpose of the matrix or vector.

[0060] Step (7) calculates the error anomaly detection index vector Q, the dynamic anomaly detection index vector G and the static anomaly detection index vector H according to the formulas Q = diag{EE T}, G = diag{SS T} and H = diag{UU T} respectively, and then sets the error anomaly control limit The dynamic anomaly control limit And the static anomaly control limit Correspond to the maximum data in Q, G and H respectively; wherein, diag{} represents the operation of converting the diagonal elements in the matrix into a column vector, and the superscript T represents the transpose of the matrix or vector.

[0061] Step (8) calculates the comprehensive anomaly detection index vector Φ according to the formula And then determines the corresponding control upper limit According to the data in Φ by using the three-sigma rule; wherein, the three-sigma rule is to first calculate the mean value b and the standard deviation c of all data in Φ, and then set

[0062] Step (9) retains the parameters for online real-time anomaly detection, and then repeatedly executes steps (9.1) to (9.4) as shown below, so as to implement online anomaly detection on a converter; wherein, the retained parameters include: the mean values μ1, μ2, …, μ 13 And the standard deviations δ1, δ2, …, δ 13 In step (2), the dynamic transformation matrix P1 and the static transformation matrix P2 in step (5), the regression coefficient matrix B in step (6), the error anomaly control limit The dynamic anomaly control limit And the static anomaly control limit In step (7), and the control upper limit

[0063] Step (9.1) collects sample data of 13 measurement variables at the latest sampling time t of a section converter, and arranges the sample data in the order of the fixed measurement variables in step (1) to form a 1x13 dimensional data vector x t Then, the average value μ1, μ2, …, μ 13 and the standard deviation δ1, δ2, …, δ 13 of the data vector x t are calculated.The standardization processing is performed on the data vector x

[0064] In the above step (9.1), the average value vector μ = [μ1, μ2, …, μ 13 ] composed of μ1, μ2, …, μ 13 and the diagonal matrix θ composed of δ1, δ2, …, δ 13 are calculated according to the formula The standardization processing is performed on the data vector x t , so that the row vector x

[0065] Step (9.2) calculates the 0th order score vector s0, the 1st order score vector s1, the 2nd order score vector s2 and the 3rd order score vector s3 according to the formula respectively, and then calculates the static score vector u t at the latest sampling time t according to the formula ; wherein, represents the row vector corresponding to the data vector x t-d composed of the data collected at the d sampling time points before t after the standardization processing.

[0066] When d = 0, represents When d = 1, represents the data vector composed of the data collected at the 1 sampling time point before t When d = 2, represents the data vector composed of the data collected at the 2 sampling time points before t When d = 3, represents the data vector composed of the data collected at the 3 sampling time points before t

[0067] Step (9.3) combines s1, s2, …, s D into an input vector , and then calculates the error vector e t according to the formula

[0068] Step (9.4) calculates the comprehensive anomaly detection index at the latest sampling time t according to the formula ​ Then, it is determined whether it is greater than If yes, the primary reformer operation is abnormal; if no, the primary reformer operation is normal.

[0069] As a preferred embodiment, when setting the time series order D in step (3), the time series order D to be set can be determined according to the following steps (3.1) to (3.4).

[0070] Step (3.1): setting α=0, D=2 and the threshold ε equal to 0.1.

[0071] Step (3.2): performing the implementation process of the 0-order time series matrix 1-order time series matrix 2-order time series matrix to D-order time series matrix in step (3).

[0072] Step (3.3): after obtaining the diagonal matrix A by performing steps (4.1) to (4.2), setting β equal to the sum of the diagonal elements in A.

[0073] Step (3.4): determining whether β-α is greater than ε; if yes, setting D=D+1 and returning to step (3.2); if no, obtaining the final value of the time series order D in step (3).

[0074] When setting the time series characteristic parameter K in step (5), the time series characteristic parameter K to be set is preferably determined according to the following steps (5.1) to (5.2).

[0075] Step (5.1): sequentially recording the elements on the diagonal line in the diagonal matrix A as a1, a2, …, a F and then setting f=1; wherein the subscript number F is equal to the number of row vectors in the diagonal matrix A.

[0076] Step (5.2): calculating the cumulative probability η f =(a1+a2+…+a f ) / (a1+a2+…+a F ) according to η f and then determining whether η f is less than 0.85; if yes, setting f=f+1 and repeating step (5.2); if no, setting the time series characteristic parameter K=f and completing the setting of the time series characteristic parameter K.

[0077] ​The above embodiments are only used to explain and illustrate the specific implementation of the present application, and are not intended to limit the present application. Any modification made to the present application within the spirit and protection scope of the claims of the present application falls within the protection scope of the present application.

Claims

1. A method for detecting operational anomalies in a single-stage converter based on time-series feature analysis, comprising the following steps (1) to (9): Step (1): After determining the measurement variables of a section of the converter, while the section of the converter is operating normally, simultaneously collect sample data for at least 800 sampling times to form a data matrix. ;in, Data Matrix Each row vector in the table is composed of sample data of the measured variables at the same sampling time, arranged in a fixed order of the measured variables. The first row of vectors to the last row are arranged in chronological order of sampling time; Step (2): Calculate the data matrix The average value of the first, second, through thirteenth column vectors in the middle. and standard deviation Then, use the mean and standard deviation to... Each row vector in the matrix is ​​standardized, and then the standardized row vectors are combined into a matrix. ; Step (3): Set the timing order Then, the matrix The Middle The row vectors to the last row vector form a 0th-order time sequence matrix. , matrix The Middle The row vectors from the second to the last row vector form a first-order time sequence matrix. Follow this pattern until you obtain... Order-order time matrix ; Step (4): For the 0th order time series matrix 1st order time sequence matrix to Order-order time matrix Perform time series feature analysis to obtain the feature transformation matrix. ; Step (5): Set timing characteristic parameters It equals a positive integer less than 13, and then the characteristic transformation matrix is... The first column vector to the second column vector Column vectors form a dynamic transformation matrix , The first in The static transformation matrix is ​​formed by the column vectors up to the 13th column vector. Then, through the formula and The 0th order score matrix was calculated separately. 1st order score matrix to rank score matrix and static score matrix ;in, ; Step (6): Determine the 0th order score matrix using the least squares regression algorithm. and Regression model between: ; where the input matrix regression coefficient matrix and error matrix The superscript T indicates the transpose of a matrix or vector; Step (7): According to the formula , and Calculate the error anomaly detection index vector respectively Dynamic anomaly detection index vector and static anomaly detection index vector Then, set error anomaly control limits. Dynamic abnormal control limits and static anomaly control limits They correspond to equal to , and The largest data in; among them, This represents the operation of converting the diagonal elements of the matrix within the curly braces into a column vector; Step (8): According to the formula Calculate the comprehensive anomaly detection index vector Then calculate The average of all data in and standard deviation Then, set the corresponding control limit. ; Step (9): After retaining the parameters used for online real-time anomaly detection, perform online anomaly detection on a section of the converter; wherein, the retained parameters include the average value in step (2). and standard deviation The dynamic transformation matrix in step (5) and static transformation matrix The regression coefficient matrix in step (6) Error anomaly control limits in step (7) Dynamic abnormal control limits and static anomaly control limits and the upper limit of control in step (8). ; Its characteristic is that the specific implementation process of step (4) is as follows: Step (4.1): For the matrix Perform singular value decomposition to obtain Then follow the formula Calculate the orthogonal transformation matrix ;in, This represents the unitary matrix produced by singular value decomposition. It is a diagonal matrix composed of non-zero singular values; Step (4.2): Using the formula Calculate the 0th order characteristic matrix respectively 1st order characteristic matrix to eigenmatrix ; Step (4.3): According to the formula Calculate the feature correlation matrix And then Perform singular value decomposition to obtain ;in, It is a diagonal matrix composed of non-zero singular values ​​arranged in descending order of their numerical values. The unitary matrix generated by singular value decomposition; Step (4.4): According to the formula The characteristic transformation matrix is ​​calculated. .

2. The method for detecting operational anomalies in a single-stage converter based on time-series feature analysis according to claim 1, characterized in that, In step (3), the timing order is set. The specific process is as follows: Step (3.1): Settings , and threshold It equals a positive number not greater than 0.1; Step (3.2): Execute the process in step (3) to form the 0th order time sequence matrix. 1st order time sequence matrix to Order-order time matrix The implementation process; Step (3.3): Perform steps (4.1) to (4.2) to obtain the diagonal matrix. Then, set equal The sum of all elements on the diagonal; Step (3.4): Judgment Is it greater than If so, then set Then, return to step (3.2); otherwise, obtain the required timing order. .

3. A method for detecting operational anomalies in a single-stage converter based on time-series feature analysis according to claim 1 or 2, characterized in that, In step (5), the timing feature parameters are set. The specific process is as follows: Step (5.1): Divide the diagonal matrix The elements on the diagonal are denoted as follows: Then, set ; where, subscript number equal to a diagonal matrix The number of row vectors in the middle; Step (5.2): According to Calculate the cumulative probability Then, make a judgment Is it less than 0.85? If so, set... Then, repeat step (5.2); otherwise, set the timing characteristic parameters. Then, the timing characteristic parameters were processed. The settings.

4. A method for detecting operational anomalies in a single-stage converter based on time-series feature analysis according to claim 1 or 2, characterized in that, The measured variables determined in step (1) are, in order: natural gas feed flow rate (F01), fuel exhaust gas feed flow rate (F02), heat exchanger outlet pressure (P01), furnace gas pressure (P02), heat exchanger outlet temperature (T01), preheated natural gas temperature (T02), mixed gas feed temperature (T03), furnace left top temperature (T04), furnace right top temperature (T05), furnace top mixed gas temperature (T06), furnace left bottom outlet temperature (T07), furnace right bottom outlet temperature (T08), and converted gas outlet temperature (T09).

5. The method for detecting operational anomalies in a single-stage converter based on time-series feature analysis according to claim 4, characterized in that, Step (9) involves repeating steps (9.1) to (9.4) below to perform online anomaly detection on a single-stage converter: Step (9.1): At the latest sampling moment of a converter section Simultaneously, sample data of 13 measurement variables were collected and arranged in the same order as the measurement variables fixed in step (1) to form a... 3D data vector Then, using the average value and standard deviation For data vectors Standardization is performed to obtain the corresponding row vectors. ; Step (9.2): According to the formula Calculate the 0th order score vector respectively 1st order score vector to 1st order score vector Then, follow the formula. Calculate the latest sampling time static score vector ;in, Indicates in Previous Data vectors collected at each sampling time point The row vector obtained after standardization, when hour, express ; Step (9.3): Combined into a single input vector Then, according to the formula Calculate the error vector ; Step (9.4): According to the formula Calculate the latest sampling time Comprehensive anomaly detection indicators Then, make a judgment Is it greater than If yes, then the first stage of the converter is malfunctioning; if no, then the first stage of the converter is operating normally.

Images