A blast furnace ironmaking process monitoring method based on RBNCVA
Through the RBNCVA model combined with SDAE and CVA, the problems of nonlinear and noise interference in blast furnace ironmaking process are solved, efficient fault detection and monitoring are achieved, false alarm rate is reduced, and fault detection rate is improved.
Patent Information
- Application Number
- CN202211258236.6
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2022-10-13
- Publication Date
- 2025-08-12
- Estimated Expiration
- 2042-10-13
AI Technical Summary
There is nonlinear complexity, noise and outlier interference in the blast furnace ironmaking process, and existing monitoring methods are difficult to effectively monitor, resulting in high false alarm rate, low fault detection rate, and lack of real-time diagnostic methods.
The robust width nonlinear features were extracted by stacked denoising autoencoder (SDAE), and combined with typical variable analysis (CVA) and kernel density estimation methods, a typical variable analysis (RBNCVA) model with robust width nonlinear representation was constructed to reduce noise interference and improve fault detection rate.
It significantly reduces the false alarm rate during blast furnace iron smelting monitoring process, improves the fault detection rate and fault sensitivity, and realizes effective real-time monitoring of blast furnace iron smelting process.
Smart Images
Figure CN115909535B_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of blast furnace ironmaking process monitoring, and in particular to a blast furnace ironmaking process monitoring method based on RBNCVA. Background Art
[0002] The steel industry is a key pillar of modern industry and the national economy. Blast furnaces are one of the most critical components in the entire steel production process, consuming over 70% of the total energy. Abnormalities in the blast furnace ironmaking process can not only lead to decreased iron production and increased energy consumption, but can also damage equipment and even cause major safety incidents. Due to the harsh production environment, complex reactions, and lack of sufficient and effective monitoring methods in large-scale blast furnace ironmaking systems, direct diagnosis of equipment status is unreliable and difficult to meet practical needs. Therefore, real-time monitoring of the blast furnace ironmaking process is of great economic and social significance.
[0003] The difficulty of monitoring the blast furnace ironmaking process stems largely from the production characteristics of the process. The interior of a blast furnace is a nonlinear, distributed system. Complex reactions occur within the blast furnace: solid (charge), liquid (slag iron), and gas (coal gas). These reactions are non-uniform along the longitudinal, radial, and circumferential directions of the blast furnace. Furthermore, the blast furnace can operate at different operating points, with nonlinearities existing between these points. The harsh production environment results in a lack of direct measurement of the blast furnace's internal conditions. The blast furnace is a sealed, high-temperature, and high-pressure vessel. The constant falling charge from above, the heavy dust, and the excessively high temperatures below prevent conventional sensors from functioning properly, thus lacking direct measurement of internal conditions. Numerous indirect measurement data are available from the top and bottom of the blast furnace, as well as from its surroundings. However, data collected from the blast furnace site contains significant noise and outliers, which significantly interfere with monitoring models. Due to the nonlinear nature of the blast furnace ironmaking process and the noise and outliers present in the data, monitoring the process is a highly challenging task.
[0004] Currently, data-driven process monitoring methods have been widely studied in large-scale blast furnace systems, particularly multivariate analysis (MVA). Canonical variate analysis (CVA), a state-space-based MVA method, is considered suitable for monitoring dynamic processes such as blast furnace ironmaking. To simultaneously account for nonlinearity and process dynamics, the kernel-regularized canonical variate analysis (KCVA) method was developed. However, the high complexity of online computation and the difficulty in selecting kernel parameters hinder the application of KCVA in practical industrial processes. Furthermore, noise and outliers in datasets also hinder the development of blast furnace ironmaking process monitoring methods. Recently, deep learning has been applied to address some of these issues, such as deep autoencoders and deep feedforward networks. Because hot blast furnace switching can cause peak interference, denoising autoencoders (DAEs) have been used in practical ironmaking monitoring. Another robust process monitoring method uses multivariate exponentially weighted moving average principal component analysis (MEWMA-PCA) and Bayesian networks (BNs). However, existing blast furnace ironmaking process monitoring methods are still far from practical application. Summary of the Invention
[0005] In order to overcome the deficiencies of the prior art, the present invention aims to provide a robust wide range nonlinear representation canonical variate analysis method (RBNCVA) to improve the monitoring performance of the blast furnace ironmaking process.
[0006] A blast furnace ironmaking process monitoring method based on robust widthwise nonlinear representation canonical variate analysis (RBNCVA) is proposed. First, a stacked denoising autoencoder (SDAE) is used to model historical blast furnace fault data to extract robust widthwise nonlinear features. This feature helps to deal with the complex nonlinearity of the blast furnace ironmaking process and is resistant to interference from noise and outliers. Then, the canonical variate analysis (CVA) method is used to analyze the relationship between past and future feature vectors to address the dynamic nature of the blast furnace. Finally, the probability density function obtained by the kernel density estimation method is used to calculate the statistical quantity.
[0007] The blast furnace ironmaking process monitoring method based on canonical variable analysis of robust width nonlinear representation comprises the following steps:
[0008] Step 1: Historical blast furnace data x collected under normal conditions o (k)∈R ζ Add a Gaussian noise vector to reconstruct the clean input data x o (k) is the target, the stacked denoising autoencoder (SDAE) is trained with weights, and a robust nonlinear feature extraction model for the blast furnace is established. The reconstruction target is as follows:
[0009]
[0010] in is a vector where each variable is independent and has mean μ and variance Gaussian distribution, θ, are the parameters of the neural network;
[0011] Step 2: SDAE uses each hidden layer to perform abstract nonlinear mapping on the blast furnace process variables, and uses the "concat" operation to fuse the features of different levels into an aggregate feature z(k), which is expressed as follows:
[0012] z(k)=[x h1 (k), x h2 (k), x h3 (k), x r (k)]∈R l
[0013] l=l h1 +l h2 +l h3 +l r
[0014] in and Respectively represent the output vector and reconstruction vector of different levels of hidden layers at the kth sampling;
[0015] Step 3: Based on the feature extraction strategies of Steps 1 and 2, use canonical variable analysis (CVA) to further explore the dynamic characteristics of the blast furnace ironmaking process; the obtained blast furnace feature matrix z∈R l*N The Hankel matrix W is divided into past and future p ∈R ls*M and W f ∈R ls*M , where the columns of the Hankel matrix are given by M = N-2s+1; the correlation between them is then maximized by solving the following optimization objective:
[0016]
[0017] stv T v=1,r T r = 1;
[0018] Step 4: Use singular value decomposition (SVD) to calculate the solution of the optimization objective in step 3.
[0019]
[0020] where V = [v1, v2, ..., v q ] and R=[r1,r2,...,r q ] are the left and right singular matrices respectively, Contains q main singular values; thus, the projection matrices J and L that maximize the correlation between the past and future features of the blast furnace are obtained:
[0021]
[0022] Step 5: Use the projection matrices J and L obtained in step 4 to convert the past blast furnace characteristic observation vectors into q-dimensional typical correlation variables and residuals respectively. The specific forms of the state space C and residual space E are as follows:
[0023] C=JW p ; E=LW p ;
[0024] Use T 2 And Q statistic is used as monitoring statistic, which is given by:
[0025]
[0026] Among them, c i (k) and e i (k) represents the i-th element of C and E at time k respectively;
[0027] Step 6: Calculate T 2 and the control limits of the Q statistic, the estimate of the probability density function is defined as follows:
[0028]
[0029] where K(·) is the kernel function, h represents the bandwidth, and the optimal bandwidth is roughly estimated by minimizing the approximate mean integrated squared error:
[0030] h opt =1.06σM -1 / 5
[0031] Where σ is the standard deviation, given the confidence level α, the monitoring statistics (T 2 The control limits for Q and Q are given by:
[0032]
[0033] Step 7: For the real-time monitoring part, first collect real-time data and perform standardization processing, and then construct the real-time statistics T through steps 2 to 6. 2 and Q, and finally judge the process operation status by comparing whether the statistics are less than the control limit. If T 2 If either Q or Q is lower than the control limit, the process is considered abnormal and needs to be stopped for inspection; otherwise, the process is considered normal and no maintenance is required.
[0034] The present invention has the beneficial effects:
[0035] The present invention proposes a RBNCVA method for monitoring the blast furnace ironmaking process. A new SDAE-based feature extraction strategy is proposed to extract robust generalized nonlinear features, which can effectively deal with the complex nonlinearities existing in blast furnace data and reduce the interference of noise and outliers in the data. The canonical variate analysis (CVA) method is then used to analyze the relationship between past and future feature vectors, extract effective monitoring statistics, and help the model deal with the dynamics existing in the blast furnace process. The present invention reduces the false alarm rate in the blast furnace ironmaking monitoring process and significantly improves the fault detection rate and fault sensitivity. BRIEF DESCRIPTION OF THE DRAWINGS
[0036] Figure 1 Shown is a schematic diagram of offline modeling of the blast furnace ironmaking process monitoring method based on RBNCVA of the present invention. DETAILED DESCRIPTION
[0037] In order to more clearly describe the technical content of the present invention, the specific implementation method is further described below with reference to the accompanying drawings.
[0038] like Figure 1 The figure shows a schematic diagram of the blast furnace ironmaking process monitoring route based on RBNCVA of the present invention, and the steps are as follows: Step 1: In the blast furnace ironmaking process, the physical and chemical reactions are complex, and there is a coupling relationship between multiple processes. Such a complex process leads to a common nonlinear relationship between different variables, so it is considered to use an autoencoder (AE) to process nonlinear data; in addition, the data collected from the blast furnace site contains large noise and outliers, which will cause great interference to the monitoring model, so a denoising autoencoder (DAE) is used to extract and synthesize robust features, and multiple DAEs are superimposed to form a deep model called a stacked denoising autoencoder (SDAE) to learn features at different levels; the historical blast furnace data x collected under normal conditions is used. o (k)∈R ζ Add a Gaussian noise vector to reconstruct the clean input data x o (k) is the target, the stacked denoising autoencoder (SDAE) is trained with weights, and a robust nonlinear feature extraction model for the blast furnace is established. The reconstruction target is as follows:
[0039]
[0040] in is a vector where each variable is independent and has mean μ and variance Gaussian distribution, θ, are the parameters of the neural network; intuitively, the model enhances robustness by trying to reconstruct normal inputs from corrupted inputs.
[0041] Step 2: SDAE uses each hidden layer to perform abstract nonlinear mapping on the blast furnace process variables. Low-level features contain more detailed information, while high-level features are more abstract and provide deeper information, but they are less sensitive to details. The “concat” operation is used to efficiently fuse features at different levels into an aggregate feature z(k) with strong representation capabilities. The expression is as follows:
[0042] z(k)=[x h1 (k), x h2 (k), x h3 (k), x r (k)]∈R l
[0043] l=l h1 +l h2 +l h3 +l r
[0044] in and Respectively represent the output vector and reconstruction vector of different levels of hidden layers at the kth sampling;
[0045] Step 3: Based on the feature extraction strategies of Steps 1 and 2, use canonical variable analysis (CVA) to further explore the dynamic characteristics of the blast furnace ironmaking process; the obtained blast furnace feature matrix Z∈R l*N The Hankel matrix W is divided into past and future p ∈R ls*M and W f ∈R ls*M , where the columns of the Hankel matrix are given by M = N-2s+1; the correlation between them is then maximized by solving the following optimization objective:
[0046]
[0047] stv T v=1,r T r=1
[0048] Step 4: Use singular value decomposition (SVD) to calculate the solution of the optimization objective in step 3.
[0049]
[0050] where V = [v1, v2, ..., v q ] and R=[r1,r2,...,r q ] are the left and right singular matrices respectively, Contains q main singular values. Thus, the projection matrices J and L that maximize the correlation between the past and future features of the blast furnace are obtained:
[0051]
[0052] Step 5: Use the projection matrices J and L obtained in step 4 to convert the past blast furnace characteristic observation vectors into q-dimensional typical correlation variables and residuals respectively. The specific forms of the state space C and residual space E are as follows:
[0053] C=JW p ; E=LW p ;
[0054] Use T 2 And Q statistic is used as monitoring statistic, which is given by:
[0055]
[0056] Among them, c i (k) and e i (k) represents the i-th element of C and E at time k respectively
[0057] Step 6: Calculate T 2 The control limits of T2 and Q statistics are obtained. Due to the non-Gaussian distribution of the state variables and residuals, the kernel density estimation (KDE) method is usually used to estimate the control limits of T2 and Q. The estimation of the probability density function is defined as follows:
[0058]
[0059] where K(·) is the kernel function, h represents the bandwidth, and the optimal bandwidth is roughly estimated by minimizing the approximate mean integrated squared error:
[0060] h opt =1.06σM -1 / 5
[0061] Where σ is the standard deviation, given the confidence level α, the monitoring statistics (T 2 The control limits for Q and Q are given by:
[0062]
[0063] Step 7: For the real-time monitoring part, first collect real-time data and perform standardization processing, and then construct the real-time statistics T through steps 2 to 6. 2 and Q, and finally judge the process operation status by comparing whether the statistics are less than the control limit. If T 2If either Q or Q is lower than the control limit, the process is considered abnormal and needs to be stopped for inspection; otherwise, the process is considered normal and no maintenance is required.
[0064] The above-described embodiments merely illustrate several implementations of the present invention, and while the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the invention. It should be noted that a person skilled in the art would be able to make numerous modifications and improvements without departing from the spirit of the present invention, all of which fall within the scope of protection of the present invention. Therefore, the scope of protection of the present invention shall be determined by the appended claims.
Claims
1. A method for monitoring a blast furnace ironmaking process based on robust wideband nonlinear representation canonical variate analysis (RBNCVA), characterized in that: Firstly, a stacked denoising autoencoder (SDAE) is used to model historical blast furnace fault data and extract robust wide nonlinear features. This feature helps to handle the complex nonlinearity of the blast furnace ironmaking process and is resistant to interference from noise and outliers. Then, the canonical variate analysis (CVA) method is used to analyze the relationship between past and future eigenvectors to address the dynamics of the blast furnace. Finally, the statistics are calculated using the probability density function obtained by the kernel density estimation method. Here are the steps: Step 1: Historical blast furnace data x collected under normal conditions o (k)∈R ζ Add a Gaussian noise vector to reconstruct the clean input data x o (k) is the target, the weights of the stacked denoising autoencoder are trained, and a robust nonlinear feature extraction model for the blast furnace is established. The reconstruction target is as follows: in is a vector where each variable is independent and has mean μ and variance Gaussian distribution, θ, are the parameters of the neural network; Step 2: The stacked denoising autoencoder uses each hidden layer to perform abstract nonlinear mapping on the blast furnace process variables, and uses the "concat" operation to fuse the features of different levels into an aggregate feature z(k), which is expressed as follows: z(k)=[x h1 (k),x h2 (k),x h3 (k),x r (k)]∈R l l＝l h1 +l h2 +l h3 +l r in and Respectively represent the output vector and reconstruction vector of different levels of hidden layers at the kth sampling; Step 3: Based on the feature extraction strategies of Steps 1 and 2, use canonical variable analysis (CVA) to further explore the dynamic characteristics of the blast furnace ironmaking process; the obtained blast furnace feature matrix Z∈R l*N The Hankel matrix W is divided into past and future p ∈R ls*M and W f ∈R ls*M , where the columns of the Hankel matrix are given by M = N-2s+1; the correlation between them is then maximized by solving the following optimization objective: s.t.v T v=1,r T r=1; Step 4: Use singular value decomposition (SVD) to calculate the solution of the optimization objective in step 3. where V=[v1,v2,…,v q ] and R=[r1,r2,…,r q ] are the left and right singular matrices respectively, Contains q main singular values; thus, the projection matrices J and L that maximize the correlation between the past and future features of the blast furnace are obtained: Step 5: Use the projection matrices J and L obtained in step 4 to convert the past blast furnace characteristic observation vectors into q-dimensional typical correlation variables and residuals respectively. The specific forms of the state space C and residual space E are as follows: c=JW p ;E=LW p ; Use T 2 And Q statistic is used as monitoring statistic, which is given by: Among them, c i (k) and e i (k) represents the i-th element of C and E at time k respectively; Step 6: Calculate T 2 and the control limits of the Q statistic, the estimate of the probability density function is defined as follows: where K(·) is the kernel function, h represents the bandwidth, and the optimal bandwidth is roughly estimated by minimizing the approximate mean integrated squared error: h opt =1.06σM -1 / 5 Where σ is the standard deviation, given the confidence level α, the monitoring statistics T 2 The control limits for and Q are given by: Step 7: For the real-time monitoring part, first collect real-time data and perform standardization processing, and then construct the real-time statistics T through steps 2 to 6. 2 and Q, and finally judge the process operation status by comparing whether the statistics are less than the control limit. If T 2 If either Q or Q is lower than the control limit, the process is considered abnormal and needs to be stopped for inspection; otherwise, the process is considered normal and no maintenance is required.