A pre-decarbonization device operation anomaly detection method based on an instant ARRG model

Abstract

The application discloses a kind of pre-decarbonization device operation anomaly detection method based on instant ARRG model, to establish adaptive AR model in real time using the online sampling data of pre-decarbonization device to generate error, then real-time error generated is used to complete operation anomaly detection task.Specifically, the method of the present application establishes an instant ARRG model in real time for each sampling time block matrix, and then calculates the error using the corresponding conversion vector and AR coefficient vector, so as to implement the operation anomaly detection of the pre-decarbonization device by the up and down fluctuation of the error.Different from the traditional method which relies on fixed time sequence relationship to describe the model, the error monitored in real time by the method of the present application is generated by the adaptive instant ARRG model, and is targeted for the sample data of each sampling time, can adapt to different characteristics of different sample data, and can always adaptively generate appropriate error for anomaly detection.

Description

TECHNICAL FIELD

[0001] The application relates to an industrial process anomaly detection method, in particular to a pre-decarbonization device operation anomaly detection method based on an instant ARRG model. BACKGROUND

[0002] Synthetic ammonia is a traditional chemical industry, which was born in the early twentieth century. In the world, ammonia is one of the most basic chemical products, which is mainly used for manufacturing nitric acid and chemical fertilizers. The production process of synthetic ammonia generally includes gas making, purification, compression and synthesis, and the production process flow includes eight processes of desulfurization, conversion, shift, decarburization, methanation, ammonia synthesis, absorption refrigeration and ammonia absorption. In the production process of synthetic ammonia, pre-removal of carbon dioxide is one of the more important processes, and the energy consumption accounts for about 10% of the total energy consumption of the ammonia plant. Therefore, the energy consumption of the carbon dioxide removal process has a great influence on the total energy consumption of the ammonia production process, and some more advanced synthetic ammonia process flows abroad have selected low-energy pre-decarbonization processes. The energy consumption of the synthetic ammonia process in China is relatively high, and on the basis of the existing pre-decarbonization device, only the stability of the operation of the entire pre-decarbonization device can be maintained to reduce energy consumption.

[0003] Therefore, real-time detection of whether the pre-decarbonization device has operation anomalies is an important technical means to maintain the stability of operation. Considering the pre-decarbonization device in the actual ammonia production, in addition to the absorption tower for absorbing carbon dioxide, there are also separators and heat exchangers, and multiple devices are coupled with each other, so that an accurate mechanism model cannot be established to implement anomaly detection. With the progress of advanced measurement instrument technology and data acquisition technology, the ammonia production process is equipped with a corresponding distributed control system (Distributed Control System, abbreviated as DCS) for data collection and storage, and it is easy to obtain real-time temperature, pressure, flow and other types of data during the operation of the pre-decarbonization device system, which lays a solid data foundation for implementing data-driven pre-decarbonization device operation anomaly detection.

[0004] Furthermore, the continuous operation of the pre-decarbonization unit dictates that its sampling data exhibits a time-series relationship. Therefore, monitoring changes in the time-series relationship within the sampling data can effectively reflect whether the pre-decarbonization unit is operating normally. In the existing technical field, time-series auto-regressive (AR) models are a classic and widely used time-series modeling technique. AR models define the corresponding time-series relationship model by describing the predictive relationship of sampling data over time. When applied to anomaly detection in pre-decarbonization units, AR models can only be built using sampling data under normal operating conditions. This type of AR model, which only reflects the time-series changes of normal data, may not have considerable sensitivity to various abnormal operating conditions. Therefore, the real-time adaptive construction of AR models and the reflection of anomalies through changes in model error are the technical challenges that need to be considered and addressed in implementing data-driven anomaly detection for pre-decarbonization units. Summary of the Invention

[0005] The main technical problem this invention aims to solve is how to utilize online sampling data from a pre-decarbonization device to establish an adaptive AR model in real time to generate errors, and then use these generated errors in real time to perform anomaly detection tasks. Specifically, to ensure the adaptive characteristics of the AR model, this invention uses an auto-regressive residual generation (ARRG) model trained in real time to generate errors that best reflect abnormal changes in the data, and utilizes the fluctuations in these errors to achieve real-time detection of whether the pre-decarbonization device is operating abnormally.

[0006] The technical solution adopted by the present invention to solve the above problems is as follows: a method for detecting abnormal operation of a pre-decarbonization unit based on an instantaneous ARRG model, comprising the following steps:

[0007] Step (1): Obtain sample data x1, x2, ..., x from the DCS historical database for N consecutive sampling times when the pre-decarbonization unit is operating under normal conditions. N Thus, the training data matrix X = [x1, x2, ..., x] is formed. N After that, the training data matrix X∈R is then processed. m×N The row vectors of each row are standardized to obtain the reference data matrix. Among them, R m×N Let R represent an m×N dimensional real matrix, and let R represent the set of real numbers.

[0008] Step (2): Using the formula Constructing a time-series block matrix where D is an integer between 2 and 6, n = N - D, subscript i ∈ {1, 2, …, n}, and superscript T represents the transpose of a matrix or a vector, represents the ith time series block matrix, R (D+1)×m represents a real matrix of (D + 1) × m dimensions, represents the column vector in the D + i-th column of R represents the column vector in the D + i - 1-th column of R represents the column vector in the i-th column of R

[0009] Step (3): Take the n time series block matrices in turn as the instantaneous block matrices respectively, and establish the instantaneous ARRG model respectively, thereby obtaining the corresponding errors e1, e2, …, en. n The specific implementation process is shown in steps (3.1) to (3.3).

[0010] Step (3.1): Initialize i = 1, and then set the instantaneous block matrix

[0011] Step (3.2): Establish the corresponding instantaneous ARRG model using the instantaneous block matrix , thereby obtaining the corresponding conversion vector w t ∈ R m×1 and the AR coefficient vector β t ∈ R 1×(D+1) The specific implementation process is shown in steps (A) to (D).

[0012] Step (A): Initialize w t to be any m × 1-dimensional real vector.

[0013] Step (B): Calculate the matrix and the matrix , and then solve the generalized eigenvalue problem G w = λL w v for the eigenvector v corresponding to the maximum eigenvalue λ, and then calculate the AR coefficient vector β t according to the formula β T = v t / v(1); where v(1) represents the first element in the eigenvector v, represents the ith time series block matrix.

[0014] Step (C): Calculate the matrix and the matrix , and then solve the generalized eigenvalue problem G β=λL β After finding the eigenvector u corresponding to the largest eigenvalue λ in u, then according to the formula w t =u / ||u|| The transformation vector w is calculated. t ; where ||u|| represents the length of u.

[0015] Step (D): Determine w t Whether it converges, and the criterion for convergence is w. t If the elements in the middle no longer change, then return to step (B); if so, then obtain the corresponding transformation vector w. t ∈R m×1 and AR coefficient vector β t ∈R 1×(D+1) Among them, R m×1 Let R represent an m×1 dimensional real vector. 1×(D+1) It represents a 1×(D+1) dimensional real vector.

[0016] Step (3.3): According to the formula The error e is calculated. i Then, check if i is less than n; if so, set i = i + 1, and then set... And return to step (3.2); otherwise, obtain e1, e2, ..., e n .

[0017] Step (4): Calculate e1, e2, ..., e n The average value μ e and standard deviation δ e Then, calculate the upper control limit C. 上 =μ e +3.5δ e and control lower limit C 下 =μ e -3.5δ e .

[0018] It should be noted that the implementation process of establishing the instant ARRG model shown in steps (A) to (D) aims to determine the optimal combination w. t ∈R m×1 and Thus, the instantaneous block matrix Error of AR model with D+1 row vectors While maximizing, ensure that the n time-series block matrices are within the same AR model. The sum of squares of the errors is minimized, that is:

[0019]

[0020] In the above formula, It is by The matrix formed by the row vectors from row 2 to row D+1. It is by The matrix formed by the row vectors from row 2 to row D+1. express The row vector in the first row of the array. express The row vector in the first row of the array.

[0021] If settings Then equation ① above can be equivalently transformed into the objective function shown below:

[0022]

[0023] The maximization problem in equation ② above can be solved using the classic Lagrange multiplier method, i.e., by constructing the Lagrange function J as shown below:

[0024]

[0025] Set J relative to w respectively t and Once the partial derivatives are equal to 0, we can obtain the generalized eigenvalue problem as shown below:

[0026]

[0027] Depend on It can be seen that β t The first element is equal to 1. Therefore, in step (B) above, the generalized eigenvalue problem G is solved. w =λL w After v, the first element v(1) of the feature vector needs to be changed to 1, which is the formula β. t =v T The implementation process of / v(1).

[0028] Steps (B) and (C) above are both for solving the same type of generalized eigenvalue problem, which can be uniformly represented as the generalized eigenvalue problem shown by Gα = λLα. Since matrix L may be non-invertible, the solution to the maximum generalized eigenvalue problem should be completed according to steps (I) to (III) as shown below.

[0029] Step (1): Perform singular value decomposition on matrix L, retaining the non-zero singular values ​​to obtain the corresponding unitary matrix U and singular value diagonal matrix V, from which the matrix can be calculated.

[0030] Step (two): According to the formula Calculate matrix Then, solve the eigenvalue problem. The eigenvector p corresponding to the largest eigenvalue λ.

[0031] Step (3): Using the formula The generalized eigenvector α is calculated.

[0032] Solve the generalized eigenvalue problem G in step (B) using the implementation process of steps (I) to (III) above. w =λL w When v, simply set G=G w L = L w And α = v. Similarly, solve the generalized eigenvalue problem G in step (C). β =λL β u then only needs to set G=G β L = L β And α = u is sufficient.

[0033] Step (5): Use DCS to acquire sample data z of the pre-decarbonization device at the latest sampling time t. t ∈R m×1 Then, for x t The data in each row undergoes the same standardization process as in step (1) to obtain a data vector.

[0034] Step (6): and the data vector of its first D sampling times Forming an instantaneous block matrix Then, a corresponding instantaneous ARRG model is established to obtain the corresponding transformation vector w. t and AR coefficient vector β t The specific implementation process is shown in steps (A) to (D).

[0035] Step (7): According to the formula Calculate the error e at the latest sampling time. t Then, determine whether condition C is satisfied. 下 ≤e t ≤C 上 If yes, the pre-decarbonization device operates normally, and the process returns to step (5) to continue using the sample data from the latest sampling time for anomaly detection; if no, step (8) is executed to decide whether to trigger an anomaly alarm.

[0036] Step (8): Return to step (5) and continue to perform anomaly detection using the sample data from the latest sampling time. If the error of ξ consecutive sampling times does not meet condition C... 下 ≤e t ≤C 上If yes, an abnormal alarm is triggered; otherwise, the pre-decarbonization device is in normal operation, and the step (5) is returned to continue the abnormality detection using the sample data at the latest sampling time; wherein, ξ is equal to the quotient of the shortest time required for triggering the abnormal alarm divided by the sampling interval time.

[0037] The advantages of the method of the present application are introduced as follows by implementing the above steps.

[0038] Firstly, the method of the present application establishes the real-time ARRG model for the block matrix at each sampling time, and then calculates the error by using the corresponding conversion vector and AR coefficient vector, so as to implement the operation abnormality detection of the pre-decarbonization device by the up-and-down fluctuation of the error. Unlike the conventional method which relies on the fixed time sequence relationship to describe the model, the error monitored by the method of the present application is generated by the adaptive real-time ARRG model, which is targeted for the sample data at each sampling time, can adapt to the different sample data with different change characteristics, and can always adaptively generate a suitable error for abnormality detection. BRIEF DESCRIPTION OF DRAWINGS

[0039] Figure 1 The flowchart of the implementation of the method of the present application is shown.

[0040] Figure 2 The schematic diagram of the pre-decarbonization device and its measurement data in the ammonia production process is shown. DETAILED DESCRIPTION

[0041] The present application will be described in detail below in combination with the drawings and specific embodiments.

[0042] The present application discloses a pre-decarbonization device operation abnormality detection method based on a real-time ARRG model, which will be described below in combination with the flowchart of the implementation of the method of the present application as shown in Figure 1

[0043] Step (1): From the DCS historical database, the sample data x1, x2, …, xN of the pre-decarbonization device running in the normal state at consecutive N sampling times are obtained. N , thereby forming a training data matrix X=[x1, x2, …, xN]. N ] is obtained, and then the row vectors of each row in the training data matrix X∈R m×N are respectively subjected to standardization processing to obtain a reference data matrix

[0044] As shown in Figure 2 ​The pre-decarbonization device shown specifically includes three devices, i.e., a separator, a heat exchanger and an absorption tower. Each sample time point collects a group of sample data, which is specifically an m x 1-dimensional real number vector composed of m = 15 measurement data, and the arrangement order is: separator inlet flow M1, separator inlet temperature M2, separator internal temperature M3, separator upper part pressure M4, separator top part outlet flow M5, heat exchanger outlet gas flow M6, heat exchanger outlet gas temperature M7, absorption tower inlet gas temperature M8, absorption tower liquid level M9, absorption tower amine liquid inlet flow M10, absorption tower amine liquid inlet temperature M11, absorption tower internal temperature M12, absorption tower upper part pressure M13, absorption tower top part outlet gas flow M14 and absorption tower top part outlet gas temperature M15.

[0045] Step (2): using the formula Assemble the time sequence block matrix Wherein, the value of time sequence order D is an integer between 2 and 6, n = N-D, subscript i ∈ {1, 2, …, n}, and superscript T represents the transpose of a matrix or a vector, Indicates the i-th time sequence block matrix, R (D+1)×m Indicates a (D+1) x m-dimensional real number matrix, Indicates the column vector of the D+i-th column in R Indicates the column vector of the D+i-1-th column in R Indicates the column vector of the i-th column in R Indicates the column vector of the i-th column in R Indicates the column vector of the i-th column in R Indicates the column vector of the i-th column in R.

[0046] Step (3): sequentially take the n time sequence block matrices R As the instantaneous block matrices R After that, the instantaneous ARRG model is established respectively, so as to obtain the corresponding errors e1, e2, …, en. n The specific implementation process is shown in steps (3.1) to (3.3);

[0047] Step (4): calculate the average value μ n And the standard deviation δ e Of e1, e2, …, en e After that, calculate the upper control limit C 上 = μ e + 3.5δ e And the lower control limit C 下 = μ e - 3.5δ e .

[0048] Step (5): use DCS to obtain the sample data z t ∈ R m×1 At the latest sampling time point t of the pre-decarbonization device, and then calculate xt The data of each row is subjected to the same standardization process as in step (1) to obtain a data vector

[0049] Step (6): the data vector and the data vectors of the previous D sampling instants are combined to form an instantaneous block matrix Then, the corresponding instantaneous ARRG model is established, thereby obtaining the corresponding conversion vector w t and the AR coefficient vector β t .

[0050] Step (7): the error e t at the latest sampling instant is calculated according to the formula Then, it is determined whether the condition C 下 ≤ e t ≤ C 上 is satisfied; if yes, the pre-decarbonization device is in normal operation, and step (5) is returned to continue abnormality detection using the sample data at the latest sampling instant; if no, step (8) is executed to determine whether to trigger an abnormality alarm.

[0051] Step (8): step (5) is returned to continue abnormality detection using the sample data at the latest sampling instant, and if the errors at the continuous ξ sampling instants do not satisfy the condition C 下 ≤ e t ≤ C 上 , an abnormality alarm is triggered; otherwise, the pre-decarbonization device is in normal operation, and step (5) is returned to continue abnormality detection using the sample data at the latest sampling instant; wherein ξ is equal to the quotient of the shortest time required to trigger an abnormality alarm divided by the sampling interval time.

Claims

1. A method for detecting abnormal operation of a pre-decarbonization device based on an instant ARRG model, characterized in that, Specifically comprising the steps shown below: Step (1): from the DCS historical database, obtain sample data x1, x2, …, xN of the pre-decarburization device running in a normal state at continuous N sampling time points N , thereby forming a training data matrix X = [x1, x2, …, xN] N ; then, normalize the row vectors of each row in the training data matrix X ∈ R m×N , to obtain a reference data matrix ; wherein, R m×N represents an m × N real matrix, and R represents a real number set; Step (2): using the formula Constructing the timing block matrix where the timing order D is an integer between 2 and 6, n = N - D, the subscript i ∈ {1, 2, …, n}, and the superscript T represents the transpose of a matrix or vector, denotes the i-th timing block matrix, R (D+1)×m denotes a real matrix of (D + 1) × m dimensions, denotes the column vector in the D + i-th column of R denotes the column vector in the D + i - 1-th column of R denotes the column vector in the i-th column of R Step (3): The time series block matrix is taken as the instantaneous block matrix and the instantaneous ARRG model is established respectively, so as to obtain the corresponding error e1, e2, …, e n The specific implementation process is shown in steps (3.1) to (3.3): Step (3.1): After initializing i = 1, the instantaneous block matrix Step (3.2): Utilizing the instant block matrix The corresponding instant ARRG model is established, and the corresponding conversion vector w is obtained t ∈R m×1 and the AR coefficient vector β t ∈R 1×(D+1) ; Step (3.3): Calculate e according to the formula Step (3.4): Calculate e1, e2, …, en according to the formula i Step (3.5): If i is less than n, set i = i + 1, and then set i = i + 1, and return to step (3.2); if not, obtain e1, e2, …, en n ; Step (4): Calculate the average μ n and the standard deviation δ e of e1, e2,..., en e After that, calculate the control upper limit C 上 = μ e + 3.5δ e and the control lower limit C 下 = μ e - 3.5δ e ; Step (5): Obtain the sample data z of the pre-decarbonization device at the latest sampling time t by using the DCS t ∈R m×1 Then, the same standardization processing as in step (1) is performed on the data of each row of x t ​ Step (6): obtaining the data vector of the (i+1)th sampling moment and the data vector of the previous D sampling moments composing an instant block matrix Then, the corresponding instant ARRG model is established, and the corresponding conversion vector w is obtained t and the AR coefficient vector β t ; Step (7): according to the formula The error e of the latest sampling time is calculated t After that, it is judged whether the condition C is met 下 ≤e t ≤C 上 If yes, the pre-decarburization device is in normal operation, and the step (5) is returned to continue to implement the anomaly detection using the sample data of the latest sampling time; if no, the step (8) is executed to decide whether to trigger an abnormal alarm; Step (8): return to step (5) to continue to implement the anomaly detection using the sample data of the latest sampling time, if the errors of the continuous ξ sampling times do not satisfy the condition C 下 ≤e t ≤C 上 , then trigger the abnormal alarm; otherwise, the pre-decarburization device is in normal operation, return to step (5) to continue to implement the anomaly detection using the sample data of the latest sampling time; wherein, ξ is equal to the quotient of the shortest time required for triggering the abnormal alarm divided by the sampling interval time.

2. The method of claim 1, wherein the method is characterized by: The sample data of each sampling time is specifically an m x 1-dimensional real number vector composed of m = 15 measurement data, and the arrangement order of the 15 measurement data is as follows: separator inlet flow, separator inlet temperature, separator internal temperature, separator upper part pressure, separator top part outlet flow, heat exchanger outlet gas flow, heat exchanger outlet gas temperature, absorption tower inlet gas temperature, absorption tower liquid level, absorption tower amine liquid inlet flow, absorption tower amine liquid inlet temperature, absorption tower internal temperature, absorption tower upper part pressure, absorption tower top part outlet gas flow, and absorption tower top part outlet gas temperature.

3. The method of claim 1, wherein the method is characterized by: The step (3.2) and step (6) establish the real-time ARRG model to obtain the corresponding conversion vector w t and AR coefficient vector β t The specific implementation process is as follows: Step (A): Initialization of w t is equal to any one of the m x 1 -dimensional real vectors; Step (B): Compute the matrix and the matrix Then solve the generalized eigenvalue problem G w = λL w for the eigenvector v corresponding to the largest eigenvalue λ of the matrix t = v T / v(1) to obtain the AR coefficient vector β t ; where v(1) represents the first element in the eigenvector v, represents the ith time block matrix; Step (C): Compute the matrix and matrix After that, solve the generalized eigenvalue problem G β = λL β u, the eigenvector u corresponding to the largest eigenvalue λ of the matrix t = u / ‖u|| to obtain the transformation vector w t ; wherein ||u|| represents the length of u; Step (D): Determine w t Has convergence been achieved? If not, return to step (B); if yes, obtain the corresponding transformation vector w. t ∈R m×1 and AR coefficient vector β t ∈R 1×(D+1) Among them, R m×1 Let R represent an m×1 dimensional real vector. 1×(D+1) It represents a 1×(D+1) dimensional real vector.

4. The method of claim 3, wherein the method is characterized by: The specific implementation process of solving the generalized eigenvalue problem in the step (B) and the step (C) is as shown below: When solving the generalized eigenvalue problem G w = λL w v in step (B), first set G = G w and L = L w , then perform steps (i) to (iii) as shown below, and then obtain the eigenvector v corresponding to the largest eigenvalue by setting v = α. Step (i): Singular Value Decomposition (SVD) is performed on matrix L, and all non-zero singular values are kept to obtain the corresponding unitary matrix U and singular value diagonal matrix V, so that matrix L can be calculated as follows Step ii. Calculate the matrix Then, solve the eigenvalue problem corresponding to the largest eigenvalue λ​ Step iii): Calculate the generalized eigenvector α by the formula ​ When solving the generalized eigenvalue problem G β = λL β u in step (C), first set G = G β and L = L β , then perform steps (i) through (iii), and then obtain the eigenvector u corresponding to the largest eigenvalue by setting u = a.

Images