A Blast Furnace Ironmaking Process Monitoring Method Based on Local-DBKSSA

Abstract

The present invention discloses a blast furnace ironmaking process monitoring method based on Local-DBKSSA to improve the monitoring performance of the blast furnace ironmaking process, including two parts: offline modeling and online monitoring; the offline modeling includes regularizing the original data, constructing a time-expanded matrix, feature extraction, SP S & NP S decomposition, statistic and threshold calculation; the online monitoring includes regularizing the test samples, time-expanded samples, feature reconstruction, SP S partial construction, local statistic, process state judgment. The present invention constructs dynamic wide non-linear features based on time shift and multi-core projection to explore process features from more perspectives. Subsequently, the above features are integrated into the steady-state subspace analysis (SSA) to accurately estimate the steady-state projection according to time-varying data. And, in order to reduce the influence of large fluctuations in the process and improve the fault detection ability, the present invention further proposes a statistic based on a local strategy.

Description

Technical Field

[0001] The present invention relates to the field of blast furnace ironmaking process monitoring, and specifically refers to a blast furnace ironmaking process monitoring method based on Local-DBKSSA (Local Dynamic Broad Kernel Steady State Subspace Analysis). Background Art

[0002] The blast furnace ironmaking process (BFIP) is the core front-end production link of modern steel manufacturing. Ensuring the stability of BFIP is also a prerequisite for enterprise operation. When anomalies occur, serious situations such as increased raw material consumption, substandard smelting quality, and increased equipment failures will occur, resulting in significant resource losses and even serious production accidents and casualties. Therefore, achieving more accurate and comprehensive process monitoring is an important means to ensure the safe and efficient operation of BFIP and is also one of the most concerned issues for the industry and scholars. Currently, the reason hindering the development of BFIP fault diagnosis is that it is a complex distributed system with difficult-to-explore non-linear time-varying dynamics, coupled with unknown physical and chemical reactions inside the blast furnace, making it difficult for researchers to accurately and completely describe its working principle with a mechanism model. However, with the development of computer and information communication technologies, data-based ironmaking process monitoring methods have come into the view of researchers and gradually attracted everyone's interest.

[0003] However, due to the extremely complex characteristics of BFIP, especially the non-stationary characteristics, its process monitoring has become an extremely challenging task, and only a few scholars are willing to try. In existing papers, adaptive models, cointegration analysis (CA) models, and steady state subspace decomposition models are regarded as feasible solutions. However, in actual BFIP, the integration orders of non-stationary variables often vary, which causes the CA-based method to lose its modeling basis and thus reduces the accuracy. Based on this understanding, the steady state subspace analysis (SSA) method based on weak stationarity is used to separate the stationary projection from the observed data, providing a practical solution. Subsequently, the exponential analytical steady state subspace analysis (EASSA) further proposed to transform the generalized eigenvalue solving problem into an exponential form to accurately obtain the stationary source and combine it with an adaptive algorithm to better adapt to the slow changes in the actual process. To explore the process dynamics, the dynamic steady state subspace analysis (DSSA-ADMM) combines the time-delay shift technology and the alternating direction method of multipliers to obtain the optimal stationary projection. However, the noise accumulation caused by the time-delay shift technology may bury useful information.

[0004] At the same time, based on our understanding of the existing technologies, when the above methods are applied to BFIP, due to the lack of consideration of multi-perspective complex dynamics and non-linearity, they all lose their ideal performance. Summary of the Invention

[0005] To this end, the present invention aims at the above-mentioned problems and proposes a blast furnace ironmaking process monitoring method based on Local-DBKSSA on the basis of BFIP characteristic analysis. For the Local-DBKSSA model, the dynamic generalized nonlinear characteristics fully explore the process nonlinearity from a multi-core perspective, construct the dynamics after the time-lag variables, and then decompose the underlying steady-state projection to mine the constant process variable relationship. In order to provide smoother and more sensitive monitoring results, a monitoring statistic based on a local strategy is constructed.

[0006] A modeling method based on local dynamic wide kernel steady-state subspace analysis (Local-DBKSSA) includes two parts: offline modeling and online monitoring;

[0007] Offline modeling includes regularizing raw data, constructing a time expansion matrix, feature extraction, SP S &NP S Decomposition, statistics and threshold calculation;

[0008] Online monitoring includes regularized test samples, time expansion samples, feature reconstruction, SP S Partial construction, local statistics, process state judgment;

[0009] Among them, in the offline modeling process, the original training data is regularized and then expanded into a time expansion matrix Z through time expansion technology. Then, multi-kernel functions are used simultaneously, and the nonlinear characteristics of the data from multiple angles are extracted and analyzed. The steady-state subspace analysis method further extracts the steady-state projections (SPs) and non-steady-state projections (NPs) in the above nonlinear characteristics. SPs are also used to construct local monitoring statistics and determine the corresponding statistical thresholds. In the online monitoring stage, the test samples are regularized and time expansion samples are constructed to extract the timing features in the test samples. Multi-perspective and steady-state features are also extracted in turn, and then the real-time statistics are calculated, and finally the process state judgment is performed.

[0010] The offline modeling part includes the following steps:

[0011] 1.1) Collect process data during normal operation, where the m-dimensional measurement vector at time t is expressed as Then, the time lag vector It is constructed as follows,

[0012]

[0013] Here the parameter q represents the number of time lags and T is the transposed sign; after collecting n samples, the time lag matrix Z is constructed,

[0014]

[0015] Here, N = n - 2q + 1;

[0016] 1.2) According to the mapping function The Gram kernel matrix K is further obtained:

[0017]

[0018] where k(.,.) is the kernel function, is the inner product in the space; the corresponding high-dimensional feature space is implicitly constructed by the kernel function structure; is obtained by centering K around the mean,

[0019]

[0020] where 1 / (n - q) constructs all elements of the matrix 1 n-q ; To further extract the main information in, referring to kernel principal component analysis (KPCA), eigenvalue decomposition is adopted here as:

[0021]

[0022] where the diagonal matrix Λ = diag(λ1, λ2,..., λ n-q ) and V = [v1, v2,..., v n-q are composed of eigenvalues and eigenvectors respectively, and the largest d (d < n - q) eigenvalues and their respective eigenvectors form the nonlinear feature where the reduced matrix is denoted as V d = [v1, v2,..., v d ;

[0023] Dynamic generalized nonlinear feature is defined by a series of features T i where f is the number of selected features;

[0024]

[0025]

[0026] 1.3) Subsequently, steady-state subspace analysis (SSA) is used to decompose into non-steady projection (NP) and steady projection (SP). Assuming that the N periods of b overflow from the average vector and covariance matrix remain unchanged for each period; The new optimization objective of the Local-DBKSSA method is derived from the following formula:

[0027]

[0028] in is the fixed projection matrix, are the mean and covariance of SP, is the average value for each period and covariance SPs and NPs pass The generalized eigenvalue decomposition of is:

[0029]

[0030]

[0031] in By the minimum d t The eigenvectors of the eigenvalues are formed, and the remaining eigenvectors form final, Can be decomposed into:

[0032]

[0033] Subsequently, a local strategy-based approach is used to perform process monitoring. First, a local SP is defined.

[0034]

[0035] where w is the length of the local statistics; the improved statistics Arranged as:

[0036]

[0037] Here the threshold is calculated by χ based on the confidence α 2 The distribution is calculated as:

[0038]

[0039] where g i =σ i / 2μ i , μ i yes The estimated mean of i yes The estimated variance of , α is the confidence level; Based on the above process, the Local-DBKSSA process monitoring model is constructed, and the corresponding statistics and thresholds are calculated.

[0040] The steps of the online monitoring are as follows:

[0041] 2.1) Given a sample of the measured variable p(t) at the t-th moment, a lag variable z(t) is formed. Assuming that the number of kernels nk has been determined, the dynamic generalized non-linear feature τ(t) based on various kernels is calculated as:

[0042]

[0043] where the centralized k1(t), …, k nk (t) is obtained by projecting the sample z(t) into different kernel spaces ;

[0044] 2.2) Then, the SP is calculated as:

[0045]

[0046] 2.3) Subsequently, a local-policy-based method can be used for process monitoring. First, a local SP is defined,

[0047]

[0048] where w is the length of the local statistic;

[0049] 2.4) The improved statistic is arranged as:

[0050]

[0051] Finally, by comparing whether the statistic is less than the threshold, the process running state is judged; if so, it is judged that the process is normal and no maintenance is required; if not, it is considered that the process has an abnormal condition and inspection needs to be stopped.

[0052] The beneficial effects of the present invention are as follows:

[0053] The present invention proposes a Local-Dynamic Broad Kernel Steady State Subspace Analysis method (Local-DBKSSA) by combining fault information, which aims to decompose more consistent steady state projections by considering multiple dynamic non-linear perspectives and construct local statistical indicators to improve the monitoring ability; a method for constructing a statistic based on a local policy is developed, and a higher fault monitoring rate and a lower false alarm rate are obtained. Description of the Drawings

[0054] Figure 1 is a schematic flow chart of the present invention. Detailed Embodiments

[0055] In order to more clearly and completely describe the technical solution of the present invention, the present invention will be further described below with reference to the drawings and embodiments.

[0056] As Figure 1 shown, a modeling method based on Local-DBKSSA (Local-Dynamic Broad Kernel Steady Subspace Analysis) of the present invention includes the following steps:

[0057] (1) For offline modeling, various sensors (flow sensors, temperature sensors, concentration sensors, etc.) need to be installed at various positions in the process, and the corresponding data is stored in a database. Subsequently, based on the data collected in the database, the process data during normal operation, where the m-dimensional measurement vector at the t-th moment can be expressed as Subsequently, the time-lagged vector can be constructed in the following manner,

[0058]

[0059] where the parameter q represents the number of time lags, and T is the transpose symbol. When n samples are collected, the time-lagged matrix Z can be constructed,

[0060]

[0061] where N = n - 2q + 1.

[0062] (2) To explore non-linearity, non-linear data can be projected onto a high-dimensional feature space suitable for linear methods The kernel method has been widely used in machine learning. According to the mapping function the Gram kernel matrix K is further obtained:

[0063]

[0064] where k(.,.) is the kernel function, is the inner product in the space. Therefore, the corresponding high-dimensional feature space is implicitly constructed by the kernel function structure. can be obtained from K centered at the mean,

[0065]

[0066] where 1 / (n - q) constructs all elements of the matrix 1 n-q . To further extract the main information in , referring to kernel principal component analysis (KPCA), eigenvalue decomposition is adopted here as:

[0067]

[0068] where the diagonal matrix Λ = diag(λ1, λ2,..., λ n-q) and V = [v1, v2, …, v n-q are respectively composed of eigenvalues and eigenvectors. The largest d (d < n - q) eigenvalues and their respective eigenvectors form the non - linear feature where the reduced matrix is denoted as V d = [v1, v2, …, v d .

[0069] However, existing literature shows that a single kernel cannot exhibit good interpolation and extrapolation capabilities. Therefore, the present invention adopts the idea of simultaneously using multiple kernel functions and complementing each other according to their respective advantages and disadvantages. Here, different kernel functions are applied to obtain various feature projections T i , i = 1, 2, …, nk. The projection differences under different kernel functions also indicate that the joint construction of multiple kernels will provide the ability to analyze the non - linearity of data from multiple perspectives. Therefore, the dynamic generalized non - linear feature is defined by a series of features T i .

[0070]

[0071] where f is the number of selected features. Since the dynamics and non - linearity of BFIP data have been fully considered, it is assumed that the dynamic non - linear feature has no autocorrelation and linearity.

[0072] (3) Subsequently, steady - state subspace analysis (SSA) is used to decompose into non - steady - state projection (NP) and steady - state projection (SP). Assume that the N time periods of b overflow from the average vector and covariance matrix which remain unchanged for each time period. The new optimization objective of the Local - DBKSSA method can be derived from the following formula:

[0073]

[0074] where is the fixed projection matrix, is the mean and covariance of SP, is the average value of each time period and covariance of SPs and of NPs can be obtained through the generalized eigenvalue decomposition of :

[0075]

[0076]

[0077] wherein is composed of eigenvectors corresponding to the smallest d t eigenvalues, and the remaining eigenvectors form Finally, can be decomposed into:

[0078]

[0079] Subsequently, a local-policy-based method can be used for process monitoring. First, a local SP is defined,

[0080]

[0081] where w is the length of the local statistic. Thus, the improved statistic is arranged as:

[0082]

[0083] where the threshold can be calculated as: based on the chi 2 distribution with confidence level α

[0084]

[0085] where g i = σ i / 2μ i , μ i is the estimated mean of, σ i is the estimated variance of, and α is the confidence level. Based on the above process, the Local-DBKSSA process monitoring model is constructed, and the corresponding statistic and threshold are calculated and obtained.

[0086] (4) For online testing, given a sample of the measured variable p(t) at the t-th moment, the lag variable z(t) is formed. Assuming that the number of kernels nk has been determined, the dynamic generalized nonlinear feature τ(t) based on various kernels can be calculated as:

[0087]

[0088] where the centralized k1(t), …, k nk (t) are obtained by projecting the sample z(t) into different kernel spaces obtained.

[0089] (5) Then, the SP can be calculated as:

[0090]

[0091] (6) Subsequently, a local-strategy-based method can be used for process monitoring. First, a local SP is defined,

[0092]

[0093] where w is the length of the local statistic.

[0094] (7) Improved statistic is arranged as:

[0095]

[0096] Finally, by comparing whether the statistic is less than the threshold, the process running state is judged. If so, it is judged that the process is normal and no maintenance is required; if not, it is considered that an abnormal condition has occurred in the process and inspection needs to be stopped.

[0097] The above description is not intended to limit the present invention. For those skilled in the art, the present invention can have various changes and modifications. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included within the protection scope of the present invention.

Claims

1. A blast furnace ironmaking process monitoring method based on Local-DBKSSA, characterized in that It includes two parts: offline modeling and online monitoring; Offline modeling includes regularizing the original data, constructing a time expansion matrix, feature extraction, SP S &NP S decomposition, statistic and threshold calculation; Online monitoring includes regularized test samples, time-expanded samples, feature reconstruction, SP S Partial structure, local statistic, process status judgment; Among them, in the offline modeling process, after the original training data is regularized, it is extended into a time-expanded matrix Z through time expansion technology. Subsequently, multiple kernel functions are used simultaneously, and the data nonlinear features from multiple angles are analyzed through feature extraction. The steady-state subspace analysis method further extracts the steady-state projections (SPs) and non-steady-state projections (NPs) in the above-mentioned nonlinear features. The SPs are also used to construct local monitoring statistics, and the corresponding statistic thresholds are determined; in the online monitoring stage, the regularized test samples are constructed to form time-expanded samples for extracting the time series features in the test samples. The multi-perspective and steady-state features are also extracted in sequence. Subsequently, the real-time statistics are calculated, and finally the process state is judged.

2. The method according to claim 1, wherein The offline modeling part includes the following steps: 1.1) Collect process data during normal operation, where the m-dimensional measurement vector at the t-th moment is expressed as Subsequently, the time-lagged vector is constructed in the following manner Here, the parameter q represents the number of time lags, and T is the transpose symbol; after collecting n samples, the time-lagged matrix Z is constructed. 1.2) According to the mapping function Further obtain the Gram kernel matrix K: where k(.,.) is a kernel function, is the inner product in space; the corresponding high-dimensional feature space is implicitly constructed by the kernel function structure; obtained by K centered at the mean, where 1 / (n - q) constructs matrix 1 n-q for all elements; in order to further extract the main information in, referring to kernel principal component analysis (KPCA), here eigenvalue decomposition is adopted as: where the diagonal matrix Λ = diag(λ1, λ2, …, λ n-q ) and V = [v1, v2, …, v n-q are composed of eigenvalues and eigenvectors respectively, d < n - q, and the largest d eigenvalues and their respective eigenvectors form the non-linear features where the reduced matrix is denoted as V d = [v1, v2, …, v d ; Dynamic generalized non-linear feature Defined by a series of features T i Define where f is the number of selected features; 1.3) Subsequently, steady-state subspace analysis (SSA) is used to decompose it into NP and SP. Assuming that the N b periods of the average vector and the covariance matrix remain unchanged for each period; the new optimization objective of the Local-DBKSSA method is derived from the following formula: where is the fixed projection matrix, is the mean and covariance of the SPs, is the average value of each period and covariance of the SPs and of the NPs are obtained by generalized eigenvalue decomposition of: Among them Composed of eigenvectors corresponding to the smallest d t eigenvalues, and the remaining eigenvectors form Finally Can be decomposed into: Subsequently, a local-policy-based method is used for process monitoring. First, a local SP is defined. where w is the length of the local statistic; the improved statistic is arranged to be: Here, the threshold is calculated as χ based on the confidence level α 2 distribution as follows: where g i = σ i / 2μ i , μ i is the estimated mean, σ i is the estimated variance, and α is the confidence level; Based on the above process, the Local-DBKSSA process monitoring model is constructed, and the corresponding statistics and thresholds are calculated and obtained.

3. The method according to claim 1, characterized in that The steps of the online monitoring are as follows: 2.1) Given the sample of the measurement vector p(t) at the t-th moment, the lag vector z(t) is formed. Assuming that the number of kernels nk has been determined, the dynamic generalized nonlinear feature τ(t) based on various kernels is calculated as: Among them, the centralized k1(t), …, k nk (t) is obtained by projecting the sample z(t) into different kernel spaces ; 2.2) Then, the SP is calculated and obtained: 2.3) Subsequently, a local-policy-based method can be used for process monitoring. First, a local SP is defined. where w is the length of the local statistic; 2.4) Improved statistic Arranged as: Finally, by comparing whether the statistic is less than the threshold, the process running state is judged; if so, it is judged that the process is normal and no maintenance is required; If not, it is considered that an abnormal situation has occurred in the process and inspection needs to be stopped.

Images