SFENOSA-based blast furnace ironmaking process monitoring method related to KPI

Abstract

The invention relates to the field of blast furnace ironmaking process monitoring and fault diagnosis, and aims to provide an SFENOSA-based blast furnace ironmaking process monitoring method related to KPI (Key Performance Indicator). The method comprises the following steps: firstly, extracting low-speed characteristics of monitoring variables through an SFA model, and then establishing an ENOSA model to construct monitoring statistics and control limits; due to the fact that the problems of overfitting and multiple collinearity in process variables are solved through ENOSA, process data and KPI data are divided into three orthogonal subspaces, and therefore independent monitoring of KPI related components and KPI unrelated components is achieved. According to the SFENOSA method provided by the invention, low-speed features extracted from process variables are taken as input, the dynamic nature of process data is overcome, and the modeling performance and the robustness are improved by utilizing an EN method; due to the fact that the dynamic nature, overfitting and multicollinearity problems of the blast furnace ironmaking process are considered, the process monitoring and fault detection efficiency can be effectively improved.

Description

Technical Field

[0001] The present invention relates to the field of blast furnace ironmaking process monitoring and fault diagnosis, and in particular to a KPI-related process monitoring method based on SFA and ENOSA (the combination of the two is referred to as SFENOSA in this application). Background Art

[0002] Blast furnace ironmaking is the process of reducing iron ore to pig iron. It mainly removes oxygen from iron ore by using the high temperature generated by the combustion of coke and reducing agents (such as carbon monoxide and hydrogen) in the blast furnace to generate liquid pig iron and slag. Molten iron is the final product of the blast furnace ironmaking process. The quality of molten iron has a great impact on the quality of subsequent steel production and even the energy consumption of the entire metallurgical production. In order to ensure the safe operation of the production process and the quality of the output products, production personnel will pay more attention to quality indicators, namely the key performance indicators (KPIs) of the production process under limited energy. In the production process, abnormal operating conditions are prone to cause process failures. If it affects product quality, it is considered to be a KPI-related failure, which may cause significant economic losses. Therefore, it is crucial to explore the correlation between easy-to-measure process variables and difficult-to-measure quality variables, monitor the fluctuation of quality indicators through changes in process variables, and study the online monitoring method of KPI-related failures to ensure the smooth progress of the ironmaking process and the quality of molten iron.

[0003] At present, in the research field of blast furnace ironmaking process monitoring, common KPI monitoring methods include OSA (orthogonal subspace analysis), PCA (principal component analysis), PLS (partial least squares), KPCA-CCA (kernel principal component analysis-canonical correlation analysis), CCA (canonical correlation analysis) methods, etc. Among them, the OSA model is mainly derived from the least squares method, which may have overfitting and multicollinearity problems; while the more traditional PCA, CCA and PLS are all based on independent assumptions and do not take into account dynamic relationships.

[0004] Therefore, there is an urgent need to propose a new KPI monitoring method to solve the above problems. Summary of the invention

[0005] The technical problem to be solved by the present invention is to overcome the deficiencies in the prior art and provide a blast furnace ironmaking process monitoring method related to KPI based on SFENOSA.

[0006] To solve the technical problem, the solution of the present invention is:

[0007] A KPI-related blast furnace ironmaking process monitoring method based on SFENOSA is provided, including extracting slow features from variables in the blast furnace ironmaking process using the SFA method, and then overcoming overfitting and multicollinearity in the process variables using the ENOSA method, dividing the process data and KPI data into three orthogonal subspaces, thereby achieving independent monitoring of KPI-related components and KPI-irrelevant components; the method specifically includes an offline training stage and an online monitoring stage; wherein,

[0008] The offline training phase includes:

[0009] (1.1) Collect sufficient variable data for the blast furnace ironmaking system in normal operation and fault conditions. The variable data includes multiple process variables and multiple KPI variables. Among them, process variables include operating variables and process variables.

[0010] (1.2) The collected variable data are standardized and then a training set is constructed; the training set includes a variable matrix used as an input for the SFA model and a KPI variable matrix used as an input for the ENOSA model;

[0011] (1.3) The training set is input into the SFA model and the ENOSA model in turn. After the former extracts the slow features, the latter continues to perform elastic network orthogonal subspace analysis.

[0012] (1.4) Input the spatial analysis results into the PCA (principal component analysis) module to construct multiple monitoring statistics, and then use the kernel density estimation method to calculate the control limits of each monitoring statistic;

[0013] The online monitoring phase includes:

[0014] (2.1) Real-time collection of variable data of the blast furnace ironmaking system during operation, and then referring to the operations in the offline training phase, the variable data are processed by the SFA model, ENOSA model and PCA module in turn, and finally the monitoring statistics are constructed;

[0015] (2.2) If the monitoring statistic constructed in step (2.1) does not exceed the control limit in step (1.4), it is considered to be fault-free and the previous step is repeated;

[0016] If the monitoring statistic constructed in this step exceeds the control limit, a warning is issued and the following is executed: when the quality-independent monitoring statistic exceeds the control limit, it indicates that there is a KPI-independent fault; when the quality-related monitoring statistic exceeds the control limit, it indicates that there is a KPI-related fault.

[0017] As a preferred embodiment of the present invention, in the step (1.1), in the variable data collected: the operating variables include at least oxygen enrichment rate, oxygen enrichment flow, cold air flow, top pressure, hot air pressure, actual wind speed, hot air temperature, blast humidity, and set coal injection amount; the process variables include at least permeability index, CO content, H 2 Content, CO 2 content, blast kinetic energy, bosh gas volume, bosh gas index, theoretical combustion temperature, oxygen enrichment pressure, cold air pressure, total pressure difference, cold air temperature, top temperature, top temperature downcomer, coal injection amount; KPI variables include at least the content of silicon (Si), phosphorus (P), and sulfur (S) in the molten iron.

[0018] As a preferred solution of the present invention, in the step (1.2), the variable data under normal operation is standardized and then constructed into a training set, and the variable data under fault state is standardized and then constructed into a test set.

[0019] As a preferred embodiment of the present invention, in step (1.3), the SFA model extracts slow features according to the following method:

[0020] (a) Given an input matrix X = [x 1 x 2 … x n ]∈R n×m , including m-dimensional process variables and n samples; let the input signal be x(t)=[x 1 (t),x 2 (t),...,x m (t)] T The goal of this algorithm is to define a transformation mapping function g(x) = [g 1 (x),g 2 (x),...,g m (x)] T , so that s j (t) = g j (x); where s j (t) is the output signal of the jth dimension;

[0021] The objective function of the mapping function is defined as:

[0022]

[0023] The constraints include:

[0024] j > t =0

[0025]

[0026] in,​ is the first-order derivative of s, <s> t is the mean of s over time, j > t = 0 means minimizing the time variation of the extracted slow features represents a simplified optimization problem, It means to avoid getting zero signal;

[0027] (b) If the mapping function is linear, then Among them, w j represents the weight vector;

[0028] (c) Finally, the optimization problem of the SFA model is transformed into a generalized eigenvalue decomposition problem: AW = BWΩ; thus, the slow feature matrix is ​​extracted as X s =W T X;

[0029] Where W is the weight matrix, A and B represent the average of the modulus length of the first-order variation of the input signal and the average of the modulus length of the input signal over time, respectively; Ω is the matrix including B -1 A is a diagonal matrix composed of the singular values ​​of A; X is the input process variable matrix, and the obtained slow features are arranged in ascending order.

[0030] As a preferred embodiment of the present invention, in step (1.3), the ENOSA model performs elastic network orthogonal subspace analysis according to the following method:

[0031] (a) Given a KPI variable matrix Y = [y 1 y 2 … y n ]∈R n×r , contains n samples with r-dimensional KPI variables collected under normal working conditions;

[0032] The slow feature matrix X extracted by the SFA model s and Y are decomposed into the following bilinear terms:

[0033]

[0034] Among them, E sENOSA and F sENOSA are the residual matrices respectively;

[0035] and is the common component matrix;

[0036] and is the transformation matrix;

[0037] ​(b) Using the ENOSA model, the input matrix is ​​decomposed into three orthogonal subspaces, namely the common component subspaces and Residual subspace E sENOSA and the residual subspace F sENOSA .

[0038] As a preferred solution of the present invention, two transformation matrices are obtained by elastic network regression method, which are as follows:

[0039] For a KPI variable, let X = [x 1 ,x 1 ,…,x m ], where x i =[x 1i ,x 1i ,…,x 1i ] T (i=1,2,...,m), KPI variable is y=[y 1 ,y 2 ,...,y n ] T , then the EN regression solution is the following optimization problem:

[0040]

[0041] Among them, λ and ρ are EN penalty parameters; β i,i＝1,...,m is the regression coefficient;

[0042] For r KPI variables, we get the transformation matrix and Further conversion to obtain and

[0043] As a preferred embodiment of the present invention, in step (1.4), the monitoring statistics and control limits are calculated according to the following method:

[0044] (a) Construct a scoring matrix:

[0045]

[0046] Among them, T comx is the score matrix, P comx is the load matrix, E xf is the residual matrix;

[0047] (b) Using Hotelling's T 2 And squared prediction error (Squared prediction error) SPE as a monitoring statistic:

[0048]

[0049] Among them, t comx It is T comx The score vector, e x is the residual E xf The score vector of ;

[0050] It is T comx The covariance matrix of

[0051] (c) Using PCA module construction Monitoring statistics of SPE x , and are constructed in the same way E sENOSA and F sENOSA Monitoring statistics of SPE y , SPE E and SPE F ; The subscripts of each statistic represent each component or subspace respectively;

[0052] (d) Construct monitoring statistics SPE xy To monitor the relationship between X and Y:

[0053]

[0054] in, When y is not considered, T comx The score vector of When considering y, T comx The score vector of Refers to the common component subspace when considering y; Y is the KPI related variable.

[0055] (e) Use kernel density estimation to calculate the control limits of each monitoring statistic.

[0056] As a preferred solution of the present invention, after offline training, a total of nine monitoring statistics are obtained: SPE x , SPE E , SPE y , SPE F 、SPE xy , and calculate the corresponding control limits; among them, SPE y , SPE F Used to observe changes in KPI variables. SPE E Used to monitor changes in process variables, SPE x At the same time, it reflects the changes of KPI related variables and process variables. xy Used to detect changes in the relationship between X and Y;

[0057] In step (2.2) of the online monitoring phase:

[0058] (a) Compare the monitoring statistics calculated in step (2.2) with the control limits calculated in the offline phase, SPE y , SPE F When the control limit is exceeded, the KPI variable is considered available; when the KPI variable is available, the fault is considered to be related to the KPI;

[0059] (b) When monitoring KPI-related failures, select SPE x , SPE y , SPE F 、SPE xy As a monitoring indicator; when monitoring KPIs irrelevant to faults, select SPE x , SPE E as a monitoring indicator.

[0060] The present invention further provides a computer device, comprising: at least one processor, and a memory communicatively connected to the at least one processor, wherein the memory stores instructions executed by the at least one processor, and the instructions are executed by the at least one processor so that the at least one processor executes the aforementioned SFENOSA-based KPI-related blast furnace ironmaking process monitoring method.

[0061] The present invention further provides a computer-readable storage medium, which stores computer instructions, and the computer instructions are used to enable the computer to execute the aforementioned blast furnace ironmaking process monitoring method based on SFENOSA and related to KPI.

[0062] Description of the invention principle:

[0063] 1. Slow feature analysis (SFA) is usually used to overcome process dynamics. In the present invention, SFA is used to extract slow features to overcome the dynamics of data in the process, so as to improve monitoring efficiency.

[0064] Elastic Net (EN) is a penalized regression model that combines the advantages of Ridge Regression and Lasso Regression (Least Absolute Shrinkage and Selection Operator, LASSO). Elastic Net has significant advantages in dealing with high-dimensional data, strong feature correlation, and problems with a large number of redundant features.

[0065] The orthonormal subspace analysis (OSA) model is mainly derived from the least squares method, which may have overfitting and multicollinearity problems. Therefore, this study uses the elastic network method to add the OSA model to improve modeling performance and enhance robustness.

[0066] In the current KPI detection technology for the production process, there is no report on the organic combination of the three.

[0067] 2. Aiming at the KPI-related faults and KPI-unrelated faults that occur in the blast furnace ironmaking process, as well as the dynamic characteristics of the operation process, the present invention innovatively proposes a KPI-related process monitoring method based on SFA and ENOSA. First, the slow characteristics of the monitoring variables are extracted through the SFA model, and then the Elastic Net Orthonormal Subspace Analysis (ENOSA) model is established to construct monitoring statistics and control limits.

[0068] 3. The ironmaking process often exhibits dynamics, so the present invention uses the slow feature analysis (SFA) method to extract slow features. Furthermore, through the elastic network orthogonal subspace analysis (ENOSA), the overfitting and multicollinearity problems in the process variables are overcome, and the process data and KPI data are divided into three orthogonal subspaces, so as to achieve independent monitoring of KPI-related components and KPI-iron components. The SFENOSA method proposed in the present invention uses the slow features extracted from the process variables as input, overcomes the dynamics of the process data, and uses the EN method to improve the modeling performance and enhance the robustness; because the dynamics of the blast furnace ironmaking process and the problems of overfitting and multicollinearity are taken into account, the efficiency of process monitoring and fault detection can be effectively improved.

[0069] Compared with the prior art, the present invention has the following beneficial effects:

[0070] 1. The KPI-related process monitoring method based on SFENOSA proposed in the present invention obtains slow features by using the SFA model, taking into account the dynamics of blast furnace ironmaking process data.

[0071] 2. The method proposed in the present invention obtains multiple monitoring statistics by establishing an ENOSA model, and can independently monitor KPI-related components and KPI-irrelevant components, effectively improving the efficiency of process monitoring. BRIEF DESCRIPTION OF THE DRAWINGS

[0072] Figure 1 It is a flow chart of a specific implementation mode of the present invention. DETAILED DESCRIPTION

[0073] The specific implementation modes of the present invention are described in detail below with reference to the accompanying drawings.

[0074] Part I Implementation of the Invention

[0075] like Figure 1 As shown, the KPI-related blast furnace ironmaking process monitoring method based on SFENOSA of the present invention includes extracting slow features from the variables in the blast furnace ironmaking process by the SFA method, and then overcoming the overfitting and multicollinearity in the process variables by the ENOSA method, dividing the process data and KPI data into three orthogonal subspaces, thereby realizing independent monitoring of KPI-related components and KPI-iron components.

[0076] The method includes an offline training phase and an online monitoring phase.

[0077] 1. Offline training phase, specifically including the following steps:

[0078] 1. Collect sufficient variable data for the blast furnace ironmaking system in normal operation and fault conditions. The variable data includes multiple process variables and multiple KPI variables. Among them, process variables include operating variables and process variables.

[0079] The operating variables include at least oxygen enrichment rate, oxygen enrichment flow, cold air flow, top pressure, hot air pressure, actual wind speed, hot air temperature, blast humidity, and set coal injection amount; the process variables include at least permeability index, CO content, H 2 Content, CO 2 content, blast kinetic energy, bosh gas volume, bosh gas index, theoretical combustion temperature, oxygen enrichment pressure, cold air pressure, total pressure difference, cold air temperature, top temperature, top temperature downcomer, coal injection amount; KPI variables include at least the content of silicon (Si), phosphorus (P), and sulfur (S) in the molten iron.

[0080] 2. Standardize the collected variable data and then construct a training set; the training set includes the variable matrix used as the input of the SFA model and the KPI variable matrix used as the input of the ENOSA model.

[0081] Among them, the variable data under normal operation is constructed as a training set after being standardized, and the variable data under fault state is constructed as a test set after being standardized.

[0082] 3. The training set is input into the SFA model and the ENOSA model in turn. After the former extracts the slow features, the latter continues to perform elastic network orthogonal subspace analysis.

[0083] (1) Establish the SFA model and extract slow features according to the following method:

[0084] (a) Given an input matrix X = [x 1 x 2 … x n ]∈R n×m , including m-dimensional process variables and n samples; let the input signal be x(t)=[x 1 (t),x 2 (t),...,x m (t)] T The goal of this algorithm is to define a transformation mapping function g(x) = [g 1 (x),g 2 (x),...,g m (x)] T , so that s j (t) = g j (x); where s j (t) is the output signal of the jth dimension;

[0085] The objective function of the mapping function is defined as:

[0086]

[0087] The constraints include:

[0088] j > t =0

[0089]

[0090] in, is the first-order derivative of s, <s> t is the mean of s over time, j > t = 0 means minimizing the time variation of the extracted slow features represents a simplified optimization problem, Indicates to avoid getting zero signal.

[0091] (b) If the mapping function is linear, then Among them, w j represents the weight vector;

[0092] (c) Finally, the optimization problem of the SFA model is transformed into a generalized eigenvalue decomposition problem: AW = BWΩ; thus, the slow feature matrix is ​​extracted as X s =W T X;

[0093] Where W is the weight matrix, A and B represent the average of the modulus length of the first-order variation of the input signal and the average of the modulus length of the input signal over time, respectively; Ω is the matrix including B -1 A is a diagonal matrix composed of the singular values ​​of A; X is the input process variable matrix, and the obtained slow features are arranged in ascending order.

[0094] (2) Establish the ENOSA model and perform elastic network orthogonal subspace analysis according to the following method:

[0095] (a) Given a KPI variable matrix Y = [y 1 y 2 …y n ]∈R n×r , contains n samples with r-dimensional KPI variables collected under normal working conditions;

[0096] The slow feature matrix X extracted by the SFA model s and Y are decomposed into the following bilinear terms:

[0097]

[0098] Among them, E sENOSA and F sENOSA are the residual matrices respectively;

[0099] and is the common component matrix;

[0100] and is the transformation matrix;

[0101] (b) Two transformation matrices are obtained by elastic network regression method, as follows: ​

[0102] For a KPI variable, let X = [x 1 ,x 1 ,...,x m ], where x i =[x 1i ,x 1i ,...,x 1i ] T (i=1,2,...,m), KPI variable is y=[y 1 ,y 2 ,...,y n ] T , then the EN regression solution is the following optimization problem:

[0103]

[0104] Among them, λ and ρ are EN penalty parameters; β i,i＝1,...,m is the regression coefficient;

[0105] Therefore, for r KPI variables, the transformation matrix can be obtained and Thus, by converting and

[0106] (c) Using the ENOSA model, the input matrix is ​​decomposed into three orthogonal subspaces, namely the common component subspaces and Residual subspace E sENOSA and the residual subspace F sENOSA .

[0107] 4. Input the spatial analysis results into the PCA (principal component analysis) module to construct multiple monitoring statistics, and then use the kernel density estimation method to calculate the control limits of each monitoring statistic;

[0108] (a) Construct a scoring matrix:

[0109]

[0110] Among them, T comx is the score matrix, P comx is the load matrix, E xf is the residual matrix;

[0111] (b) Using Hotelling's T 2 And squared prediction error (Squared prediction error) SPE as a monitoring statistic:

[0112]

[0113] Among them, t comx It is T comx The score vector, e x is the residual E xf The score vector of ;

[0114] It is T comx The covariance matrix of

[0115] (c) Using PCA module construction Monitoring statistics of SPE x , and are constructed in the same way E sENOSA and F sENOSA Monitoring statistics of SPE y , SPE E and SPE F ; The subscripts of each statistic represent each component or subspace respectively;

[0116] (d) Construct monitoring statistics SPE xy To monitor the relationship between X and Y:

[0117]

[0118] in, When y is not considered, T comx The score vector of When considering y, T comx The score vector of Refers to the common component subspace when considering y; Y is the KPI related variable.

[0119] (e) Use kernel density estimation to calculate the control limits of each monitoring statistic.

[0120] After offline training, a total of nine monitoring statistics are obtained, namely SPE x , SPE E , SPE y , SPE F 、SPE xy , and calculate the corresponding control limits; among them, SPE y , SPE F Used to observe changes in KPI variables. SPE E Used to monitor changes in process variables, SPE x At the same time, it reflects the changes of KPI related variables and process variables. xy Used to detect changes in the relationship between X and Y.

[0121] 2. The online monitoring stage includes the following steps:

[0122] 1. Collect variable data of the blast furnace ironmaking system in real time during operation, and then refer to the operations in the offline training stage to process the variable data through the SFA model, ENOSA model and PCA module in turn, and finally construct monitoring statistics;

[0123] 2. If the monitoring statistic constructed in the previous step does not exceed the control limit in step 4 of the offline phase, it is considered to be fault-free and the previous step is repeated;

[0124] If the monitoring statistic constructed in this step exceeds the control limit, a warning is issued and the following is executed: when the quality-independent monitoring statistic exceeds the control limit, it indicates that there is a KPI-independent fault; when the quality-related monitoring statistic exceeds the control limit, it indicates that there is a KPI-related fault.

[0125] In step 2 of the online monitoring phase:

[0126] (a) Compare the monitoring statistics calculated in this step with the control limits calculated in the offline phase, SPE y , SPE F When the control limit is exceeded, the KPI variable is considered available; when the KPI variable is available, the fault is considered to be related to the KPI;

[0127] (b) When monitoring KPI-related failures, select SPE x , SPE y , SPE F 、SPE xy As a monitoring indicator; when monitoring KPIs irrelevant to faults, select SPE x , SPE E as a monitoring indicator.

[0128] 3. To implement the above method, the present invention provides a suitable computer device and a computer-readable storage medium.

[0129] The computer device includes: at least one processor, and a memory communicatively connected to the at least one processor, wherein the memory stores instructions executed by the at least one processor, and the instructions are executed by the at least one processor so that the at least one processor executes a blast furnace ironmaking process monitoring method related to KPI based on SFENOSA.

[0130] The computer-readable storage medium stores computer instructions, and the computer instructions are used to enable the computer to execute a blast furnace ironmaking process monitoring method related to KPI based on SFENOSA.

[0131] Part II Contents of a Verification Experiment

[0132] The following simulation experiment is designed based on a real blast furnace ironmaking process (BFIP) data set to verify the effectiveness of the fault detection method for blast furnace steel production process based on SFA and ENOSA. The blast furnace ironmaking process in this simulation experiment is divided into four main units: blast furnace body, blast furnace top, pulverized coal injection system and hot air system. The data set comes from the blast furnace system of a steel company in Liuzhou, Guangxi.

[0133] Molten iron is the final product of blast furnace ironmaking. Its quality has a great impact on the quality of subsequent steel and the energy consumption of the entire metallurgical production. Phosphorus content [P], sulfur content [S], and silicon content [Si] are the most critical indicators for evaluating molten iron quality.

[0134] In this embodiment, 13 operating variables and 20 process variables are selected as input variables, as shown in Table 1. Three KPI variables ([Si], [P], [S]) are selected as output variables, 500 data under normal working conditions are collected as training sets, and 180 data are collected as test sets, among which the suspension failure occurs in the 71st sample.

[0135] In this example, OSA, ENOSA, PLS, KPCA-CCA, and CCA methods are used as prior art for comparison with the method of the present invention. In these prior art methods, the fault detection rate (FDR) of KPI-related faults is obtained by selecting appropriate parameters. The details are shown in Table 2.

[0136] Table 1 Variables monitored during blast furnace ironmaking

[0137]

[0138] Table 2 Fault detection rate (FDR) of various methods for KPI-related faults (%)

[0139]

[0140] It can be seen from the data in Table 1 that, compared with several other prior art methods, the method of the present invention has the highest fault detection rate for KPI-related faults. Therefore, the present invention has a more efficient monitoring capability for KPI-related processes.< / s> ​< / s>

Claims

1. A blast furnace ironmaking process monitoring method related to KPI based on SFENOSA, characterized in that: This includes using the SFA method to extract slow features from variables in the blast furnace ironmaking process, and then using the ENOSA method to overcome overfitting and multicollinearity in process variables, dividing process data and KPI data into three orthogonal subspaces, thereby achieving independent monitoring of KPI-related components and KPI-iron components; The method specifically includes an offline training phase and an online monitoring phase; wherein, The offline training phase includes: (1.1) Collect sufficient variable data for the blast furnace ironmaking system in normal operation and fault conditions. The variable data includes multiple process variables and multiple KPI variables. Among them, process variables include operating variables and process variables. (1.2) The collected variable data are standardized and then a training set is constructed; the training set includes a variable matrix used as an input for the SFA model and a KPI variable matrix used as an input for the ENOSA model; (1.3) The training set is input into the SFA model and the ENOSA model in turn. After the former extracts the slow features, the latter continues to perform elastic network orthogonal subspace analysis. (1.4) Input the spatial analysis results into the PCA module to construct multiple monitoring statistics, and then use the kernel density estimation method to calculate the control limits of each monitoring statistic; The online monitoring phase includes: (2.1) Real-time collection of variable data of the blast furnace ironmaking system during operation, and then referring to the operations in the offline training phase, the variable data are processed by the SFA model, ENOSA model and PCA module in turn, and finally the monitoring statistics are constructed; (2.2) If the monitoring statistic constructed in step (2.1) does not exceed the control limit in step (1.4), it is considered to be fault-free and the previous step is repeated; If the monitoring statistic constructed in this step exceeds the control limit, a warning is issued and the following is executed: when the quality-independent monitoring statistic exceeds the control limit, it indicates that there is a KPI-independent fault; when the quality-related monitoring statistic exceeds the control limit, it indicates that there is a KPI-related fault.

2. The method according to claim 1, characterized in that: In the step (1.1), among the variable data collected: the operating variables include at least oxygen enrichment rate, oxygen enrichment flow, cold air flow, top pressure, hot air pressure, actual wind speed, hot air temperature, blast humidity, and set coal injection amount; the process variables include at least permeability index, CO content, H2 content, CO2 content, blast kinetic energy, bosh gas volume, bosh gas index, theoretical combustion temperature, oxygen enrichment pressure, cold air pressure, total pressure difference, cold air temperature, top temperature, top temperature downcomer, and coal injection amount; the KPI variables include at least the content of silicon (Si), phosphorus (P), and sulfur (S) in molten iron.

3. The method according to claim 1, characterized in that: In the step (1.2), the variable data under normal operation is standardized and then constructed into a training set, and the variable data under fault state is standardized and then constructed into a test set.

4. The method according to claim 1, characterized in that In the step (1.3), the SFA model extracts slow features according to the following method: (a) Given an input matrix X = [x1 x2 … x n ]∈R n×m , including m-dimensional process variables and n samples; let the input signal be x(t)=[x1(t),x2(t),…,x m (t)] T The goal of this algorithm is to define a transformation mapping function g(x) = [g1(x), g2(x), ..., g m (x)] T , so that s j (t) = g j (x); where s j (t) is the output signal of the jth dimension; The objective function of the mapping function is defined as: The constraints include: <s j > t ＝0 in, is the first-order derivative of s, <s> t is the mean of s over time, j > t = 0 means minimizing the time variation of the extracted slow features represents a simplified optimization problem, It means to avoid getting zero signal;​< / s> <s> (b) If the mapping function is linear, then Among them, w j represents the weight vector; (c) Finally, the optimization problem of the SFA model is transformed into a generalized eigenvalue decomposition problem: AW = BWΩ; thus, the slow feature matrix is ​​extracted as X s =W T X; Where W is the weight matrix, A and B represent the average of the modulus length of the first-order variation of the input signal and the average of the modulus length of the input signal over time, respectively; Ω is the matrix including B -1 A is a diagonal matrix composed of the singular values ​​of A; X is the input process variable matrix, and the obtained slow features are arranged in ascending order.

5. The method according to claim 1, characterized in that In the step (1.3), the ENOSA model performs elastic network orthogonal subspace analysis according to the following method: (a) Given a KPI variable matrix Y = [y1 y2 ... y n ]∈R n×r , contains n samples with r-dimensional KPI variables collected under normal working conditions; The slow feature matrix X extracted by the SFA model s and Y are decomposed into the following bilinear terms: Among them, E sENOSA and F sENOSA are the residual matrices respectively; and is the common component matrix; and is the transformation matrix; (b) Using the ENOSA model, the input matrix is ​​decomposed into three orthogonal subspaces, namely the common component subspaces and Residual subspace E sENOSA and the residual subspace F sENOSA .

6. The method according to claim 5, characterized in that Two transformation matrices are obtained through the elastic network regression method, as follows: For a KPI variable, let X = [x1, x1, ..., x m ], where x i =[x 1i ,x 1i ,...,x 1i ] T (i=1,2,...,m), KPI variable is y=[y1,y2,...,y n ] T , then the EN regression solution is the following optimization problem: Among them, λ and ρ are EN penalty parameters; β i,i＝1,...,m is the regression coefficient; For r KPI variables, we get the transformation matrix and Further conversion to obtain and 7. The method according to claim 1, characterized in that In the step (1.4), the monitoring statistics and control limits are calculated according to the following method: (a) Construct a scoring matrix: Among them, T comx is the score matrix, P comx is the load matrix, E xf is the residual matrix; (b) Using Hotelling's T 2 And squared prediction error (SPE) as a monitoring statistic: Among them, t comx It is T comx The score vector, e x is the residual E xf The score vector of ; It is T comx The covariance matrix of (c) Using PCA module construction Monitoring statistics of SPE x , and are constructed in the same way E sENOSA and F sENOSA Monitoring statistics of and Among them, the subscripts of each statistic represent each component or subspace; (d) Construct monitoring statistics SPE xy To monitor the relationship between X and Y: in, When y is not considered, T comx The score vector of When considering y, T comx The score vector of Refers to the common component subspace when considering y; Y is the KPI related variable; (e) Use kernel density estimation to calculate the control limits of each monitoring statistic.

8. The method according to claim 1, characterized in that After offline training, a total of nine monitoring statistics are obtained: And calculate the corresponding control limits; among them, Used to observe changes in KPI variables. Used to monitor changes in process variables, At the same time, it reflects the changes of KPI related variables and process variables. xy Used to detect changes in the relationship between X and Y; In step (2.2) of the online monitoring phase: (a) Compare the monitoring statistics calculated in step (2.2) with the control limits calculated in the offline phase When the control limit is exceeded, the KPI variable is considered available; when the KPI variable is available, the fault is considered to be related to the KPI; (b) When monitoring KPI-related failures, select As a monitoring indicator; when monitoring KPIs irrelevant to faults, select as a monitoring indicator.

9. A computer device, characterized in that: include: At least one processor, and a memory communicatively connected to the at least one processor, wherein the memory stores instructions executed by the at least one processor, and the instructions are executed by the at least one processor so that the at least one processor executes the blast furnace ironmaking process monitoring method related to KPI based on SFENOSA as described in any one of claims 1 to 8.

10. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores computer instructions, and the computer instructions are used to enable the computer to execute the blast furnace ironmaking process monitoring method related to KPI based on SFENOSA according to any one of claims 1 to 8. < / s>

Images