Continuous annealing process fault detection method based on recursive kernel principal component analysis

Abstract

A fault detection method in a continuous annealing process based on a recursive kernel principal component analysis (RKPCA) is disclosed. The method includes: collecting data of the continuous annealing process including roll speed, current and tension of an entry loop (ELP); building a model using the RKPCA and updating the model, and calculating the eigenvectors {circumflex over (P)}. In the fault detection of the continuous annealing process, when the T2 statistic and SPE statistic are greater than their confidence limit, a fault is identified; on the contrary, the whole process is normal. The method mainly solves the nonlinear and time-varying problems of data, updates the model and calculates recursively the eigenvalues and eigenvectors of the training data covariance by the RKPCA. The results show that the method can not only greatly reduce false alarms, but also improve the accuracy of fault detection.

Description

BACKGROUND OF THE INVENTION

1. Field of the Invention

The present invention relates to a fault detection and diagnosis, more particular to a fault monitoring method of a continuous annealing process based on recursive kernel principal component analysis.

2. The Prior Arts

With increasing complexity of industrial processes, the requirement for reliability, availability and security is growing significantly. Fault detection and diagnosis (FDD) are becoming a major issue in industry. The actual production process has different characteristics, like linear, nonlinear, time-invariant, time-varying, etc. For the different production processes, we should use different fault monitoring methods so as to effectively monitor the fault. Continuous annealing process is a complex time-varying nonlinear process.

For nonlinear characteristics of the industrial process, some scholars have proposed a kernel principal analysis (KPCA) method. KPCA projects nonlinear data to high-dimensional feature space by nonlinear kernel function, then performs a linear PCA feature extraction in the feature space. KPCA is to perform PCA in high-dimensional feature space, which is not necessary for solving nonlinear optimization problems, and compared with other nonlinear methods it does not need to specify the number of the principal component before modeling, but KPCA method has disadvantage. KPCA is an approach based on the data covariance structure where the principal component model is time-invariant. In the actual industrial process, the mean, variance, correlation structure of process variables under normal conditions will be changed slowly due to sensor drift, equipment aging, raw material change and reduced catalyst activity, etc. Compared with the process fault the changes are slow, which belongs to the normal process operation. When the time-invariant principal component model is applied to time-varying process, it may cause false alarms. Therefore it is necessary to propose a feasible method to solve the time-varying nonlinear problems.

SUMMARY OF THE INVENTION

To solve the time-varying nonlinear problems, the invention proposes a fault monitoring method in a continuous annealing process based on recursive kernel principal component analysis to achieve the purpose of reducing false alarm rate.

The technical solution of the present invention is implemented as follows: The fault detection method in the continuous annealing process based on a recursion kernel principal component analysis (RKPCA) includes the following steps:

Step 1: Collect data and standardize the data, collecting data in the continuous annealing industrial process including: the roll speed, current and tension of an entry loop (ELP);

Step 2: Calculate the principal factors P of fault in the continuous annealing process, i.e., build an initial monitoring model of the continuous annealing process with N standardized samples in Step 1. Monitor a new sample x_new of the continuous annealing process. If it is abnormal, an alarm will be given, otherwise go to Step 3;

Where, the extracted principal factor P in the continuous annealing process is as follows:

$P=\Phi \left(X\right)\left[\begin{array}{cc}\frac{1}{h_{\Phi }}\sqrt{\frac{N-1}{N\left(N-2\right)}}& 0^T\\ -\frac{1}{h_{\Phi }}\sqrt{\frac{N-1}{N\left(N-2\right)}}B& \tilde{A}\end{array}\right]U_{\Phi }^{\prime }$ Here Φ(X) is a mapping matrix of N samples X=[x₁, x₂, . . . , x_N]. N is the sample number, the regulating factor of initial monitoring model in the continuous annealing process is

$h_{\Phi }=\frac{N-1}{N\left(N-2\right)}\sqrt{1-2B^{T}k\left(X,x_1\right)+B^{T}K\left(X\right)B},$ the correcting matrix of initial monitoring model in the continuous annealing process is

$B=\frac{1}{N-1}{1}_{N-1}+\tilde{A}\tilde{\Lambda }{\tilde{A}}^T\left(k\left(\tilde{X},x_1\right)-\frac{1}{N-1}K\left(\tilde{X}\right)1_{N-1}\right),$ k(X, x₁) indicates the inner product of X and x₁. K(X) indicates the inner product of the sample matrix. k({tilde over (X)}, x₁) is the inner product of {tilde over (X)} and x₁, {tilde over (X)} is the middle matrix, K({tilde over (X)}) indicates the inner product of the middle matrix, {tilde over (Λ)} is the eigenvalues matrix of the middle covariance matrix, U_Φ′ is the eigenvectors matrix of the process variables, 1_N-1 is the unit vector in N−1 column;Extract the transmission factor of the continuous annealing process, which is expressed as:

$\left[\begin{array}{cc}\frac{1}{h_{\Phi }}\sqrt{\frac{N-1}{N\left(N-2\right)}}& 0^T\\ -\frac{1}{h_{\Phi }}\sqrt{\frac{N-1}{N\left(N-2\right)}}B& \tilde{A}\end{array}\right]={A\left(U_{\Phi }^{\prime }\right)}^{-1}$

Step 3: When the continuous annealing process sample x_new is normal data, we use the recursive kernel principal component analysis (RKPCA) in Step 2 to update the initial monitoring model of the continuous annealing process, and calculate the principal factor {circumflex over (P)} of the fault in the updated continuous annealing process model. {circumflex over (P)} is expressed as follows:

$\hat{P}=\Phi \left(\left[\begin{array}{cc}\tilde{X}& x_{\mathrm{new}}\end{array}\right]\right)\left[\begin{array}{cc}\tilde{A}& -\frac{1}{h_{\Phi }^{\prime }}\sqrt{\frac{N-1}{N\left(N-2\right)}}{B}^{\prime }\\ 0^T& \frac{1}{h_{\Phi }^{\prime }}\sqrt{\frac{N-1}{N\left(N-2\right)}}\end{array}\right]U_{\Phi }^{″}=\Phi \left(X_{\mathrm{new}}\right)\hat{A}$ where Φ(X_new)=Φ([{tilde over (X)} x_new]) is the updated mapping matrix. The regulating factor for the updated monitoring model in the continuous annealing process is

$h_{\Phi }^{\prime }=\frac{N-1}{N\left(N-2\right)}\sqrt{1-2B^{\prime T}k\left(\tilde{X},x_{\mathrm{new}}\right)+B^{\prime T}K\left(\tilde{X}\right)B^{\prime }},$ the regulating matrix for the updating monitoring model in the continuous annealing process is

$B^{\prime }=\frac{1}{N-1}{1}_{N-1}+\tilde{A}\tilde{\Lambda }{\tilde{A}}^T\left(k\left(\tilde{X},x_{\mathrm{new}}\right)-\frac{1}{N-1}K\left(\tilde{X}\right)1_{N-1}\right),k\left(\tilde{X},x_{\mathrm{new}}\right)$ $m_{\Phi }=\frac{1}{N}\Phi \left(\left[\begin{array}{cc}{x}_1& \tilde{X}\end{array}\right]\right)1_N=\frac{1}{N}\Phi \left(x_1\right)+\frac{N-1}{N}{\tilde{m}}_{\Phi }$ $C^F=\frac{1}{N-1}\stackrel{\_}{\Phi }\left(\left[\begin{array}{cc}{x}_1& \tilde{X}\end{array}\right]\right){\stackrel{\_}{\Phi }\left(\left[\begin{array}{cc}{x}_1& \tilde{X}\end{array}\right]\right)}^T$ indicates the inner product of {tilde over (X)} and x_new.

Step 4: Detect Fault for the Continuous Annealing Process.

The fault of the continuous annealing process can be judged by using Hotelling's T² statistic and squared prediction error (SPE) statistic. When the T² statistic and SPE statistic exceed their confidence limit, a failure is identified; on the contrary, the whole process is normal, go to step 3 to continue to update the initial monitoring model of the continuous annealing process.

Step 2 describes that an initial model of the continuous annealing process using the first N samples after standardizing in Step 1 is built, and includes the following steps:

RKPCA method proposed by the invention updates recursively eigenvalues in the feature space of the sample covariance matrix. Let X=[x₁, x₂, . . . , x_N] be the sample matrix of the continuous annealing process, x₁, x₂, . . . , x_N are the samples of the continuous annealing process, N is the sample number, {tilde over (X)}=[x₂, . . . , x_N]εR^mx(N-1) is the middle matrix of the continuous annealing process, m is the number of sampling variables in the continuous annealing process, X_new=[{tilde over (X)} x_new] is the sample matrix of updating model in the continuous annealing process, x_new is the new sample of the continuous annealing process. After mapping X, {tilde over (X)} and X_new to the high-dimensional feature space, they are Φ(X), Φ({tilde over (X)}) and Φ(X_new) respectively. So the mean vector m_Φ and covariance matrix C^F of Φ(X) can be calculated

$\begin{array}{cc}{m}_{\Phi }=\frac{1}{N}\Phi \left(\left[\begin{array}{cc}{x}_1& \tilde{X}\end{array}\right]\right)1_N=\frac{1}{N}\Phi \left(x_1\right)+\frac{N-1}{N}{\tilde{m}}_{\Phi }{C}^F=\frac{1}{N-1}\stackrel{\_}{\Phi }\left(\left[\begin{array}{cc}{x}_1& \tilde{X}\end{array}\right]\right){\stackrel{\_}{\Phi }\left(\left[\begin{array}{cc}{x}_1& \tilde{X}\end{array}\right]\right)}^T& \left(1\right)\\ \begin{array}{c}=\frac{1}{N-1}\left(\Phi \left(x_1\right)-m_{\Phi }\right){\left(\Phi \left(x_1\right)-m_{\Phi }\right)}^T+\frac{1}{N-1}\sum _{i=2}^N\\ \left(\Phi \left(x_i\right)-m_{\Phi }\right){\left(\Phi \left(x_i\right)-m_{\Phi }\right)}^T\end{array}& \\ \begin{array}{c}=\frac{1}{N-1}\left[\frac{N-1}{N}\Phi \left(x_1\right)-\frac{N-1}{N}{\tilde{m}}_{\Phi }\right]\\ {\left[\frac{N-1}{N}\Phi \left(x_1\right)-\frac{N-1}{N}{\tilde{m}}_{\Phi }\right]}^T+\frac{1}{N-1}\sum _{i=2}^N\\ \left[\Phi \left(x_i\right)-{\tilde{m}}_{\Phi }+\frac{1}{N}{\tilde{m}}_{\Phi }-\frac{1}{N}\Phi \left(x_1\right)\right]\times \end{array}& \\ \begin{array}{c}\mathrm{ }{\left[\Phi \left(x_i\right)-{\tilde{m}}_{\Phi }+\frac{1}{N}{\tilde{m}}_{\Phi }-\frac{1}{N}\Phi \left(x_1\right)\right]}^T\\ =\frac{1}{N}\left(\Phi \left(x_1\right)-{\tilde{m}}_{\Phi }\right){\left(\Phi \left(x_1\right)-{\tilde{m}}_{\Phi }\right)}^T+\frac{1}{N-1}\sum _{i=2}^N\\ \left(\Phi \left(x_i\right)-{\tilde{m}}_{\Phi }\right){\left(\Phi \left(x_i\right)-{\tilde{m}}_{\Phi }\right)}^T\\ =\frac{1}{N}\left(\Phi \left(x_1\right)-{\tilde{m}}_{\Phi }\right){\left(\Phi \left(x_1\right)-{\tilde{m}}_{\Phi }\right)}^T+\frac{N-2}{N-1}{\tilde{C}}^F\end{array}& \\ \begin{array}{c}=\frac{N-2}{N-1}\left[\begin{array}{cc}\sqrt{\frac{N-1}{N\left(N-2\right)}}\left(\Phi \left(x_1\right)-{\tilde{m}}_{\Phi }\right)& \sqrt{\frac{1}{N-2}}\stackrel{\_}{\Phi }\left(\tilde{X}\right)\end{array}\right]\times \\ \left[\begin{array}{cc}\sqrt{\frac{N-1}{N\left(N-2\right)}}\left(\Phi \left(x_1\right)-{\tilde{m}}_{\Phi }\right)& \sqrt{\frac{1}{N-2}}\stackrel{\_}{\Phi }\left(\tilde{X}\right)\end{array}\right]\end{array}& \left(2\right)\end{array}$ where {tilde over (m)}_Φ and {tilde over (C)}^F represent the mean vector and covariance matrix of Φ({tilde over (X)}), respectively. Φ([x₁ {tilde over (X)}]) is the mean matrix of Φ(X), 1_N is a row vector consisting of 1 with the number of N, Φ(x_i) is the mapping vector to the high-dimensional feature space of x_i, i=1 . . . N, Φ({circumflex over (X)}) is the mean matrix of Φ({tilde over (X)});Λ and P are the eigenvalues matrix and the main factors of C^F, respectively. Λ and {tilde over (P)} are the eigenvalues matrix and the main factors of covariance matrix {tilde over (C)}^F of Φ({tilde over (X)}), respectively. Assume {tilde over (P)}=PR_Φ, R_Φ is an orthogonal rotation matrix. Due to P=Φ(X)A, {tilde over (P)}=Φ({tilde over (X)})Ã, where A=(I−(1/N)×E_N)[v₁/√{square root over (ξ₁)}, v₂/√{square root over (ξ₂)}, . . . , v_i/√{square root over (ξ_i)}], ξ_i and v_i indicate the ith eigenvalues and eigenvectors of Φ(X)^TΦ(x), respectively. Ã=(I−(1/(N−1))×E_N-1)[{tilde over (v)}₁/√{square root over (ω₁)}, {tilde over (v)}₂/√{square root over (ω₂)}, . . . , {tilde over (v)}_i/√{square root over (ω_i)}], ω_i and {tilde over (v)}_i indicate the ith eigenvalues and eigenvectors of Φ({tilde over (X)})^TΦ({tilde over (X)}), we can get P^TC^FP=Λ and {tilde over (P)}^T{tilde over (C)}^F{tilde over (P)}={tilde over (Λ)} by diagonalizing C^F and {tilde over (C)}^F, respectively. We can get [(N−1)/(N−2)]Λ−[(N−1)/(N(N−2))]g_Φg_Φ^T=R_Φ{tilde over (Λ)}R_Φ^T from Equation (2), wherein g_Φ=P^T(Φ(x₁)−{tilde over (m)}_Φ)=A^T[k(X,x₁)−(1/(N−1))K(X, {tilde over (X)})1_N-1]; Let S_Φ=[(N−1)/(N−2)]Λ−[(N−1)/(N(N−2))]g_Φg_Φ^T, {tilde over (Λ)} and R_Φ are the eigenvalues matrix and eigenvectors matrix of S_Φ, we get Equation (3) from Equation (2).

$\begin{array}{cc}\begin{array}{c}{P}^T{C}^{F}P=\frac{1}{N}{P}^T\left(\Phi \left(x_1\right)-{\tilde{m}}_{\Phi }\right){\left(\Phi \left(x_1\right)-{\tilde{m}}_{\Phi }\right)}^{T}P+\frac{N-2}{N-1}{P}^T{\tilde{C}}^{F}P\\ =\frac{1}{N}{g}_{\Phi }{g}_{\Phi }^T+\frac{N-2}{N-1}{A}^T{\Phi \left(X\right)}^T\tilde{P}\tilde{\Lambda }{\tilde{P}}^T\Phi \left(X\right)A\\ =\frac{1}{N}{g}_{\Phi }{g}_{\Phi }^T+\frac{N-2}{N-1}{A}^T{\Phi \left(X\right)}^T\Phi \left(\tilde{X}\right)\tilde{A}\tilde{\Lambda }{\tilde{A}}^T{\Phi \left(\tilde{X}\right)}^T\Phi \left(X\right)A\\ =\frac{1}{N}{g}_{\Phi }{g}_{\Phi }^T+\frac{N-2}{N-1}{A}^{T}K\left(X,\tilde{X}\right)\tilde{A}\tilde{\Lambda }{\tilde{A}}^T{K\left(X,\tilde{X}\right)}^{T}A\\ =\Lambda \end{array}& \left(3\right)\end{array}$ where K(X,{tilde over (X)}) indicates the inner product of sample matrix and middle matrix in the continuous annealing process;The singular value decomposition in Equation (2) satisfies:

$\begin{array}{cc}\sqrt{\frac{1}{N-2}}\stackrel{\_}{\Phi }\left(\tilde{X}\right)=\tilde{P}\widetilde{\sum _{\Phi }^{}}{\tilde{D}}_{\Phi }^T& \left(4\right)\end{array}$ where {tilde over (P)}=Φ({tilde over (X)})Ã is the main factor of {tilde over (C)}^F, {tilde over (Σ)}_Φ is the diagonal matrix and satisfies {tilde over (Σ)}_Φ²={tilde over (Λ)}. {tilde over (D)}_Φ is the corresponding right-singular matrix. From Equations (4) and (2), we have:

$\begin{array}{cc}\left[\sqrt{\frac{N-1}{N\left(N-2\right)}}\left(\Phi \left(x_1\right)-{\tilde{m}}_{\Phi }\right)\sqrt{\frac{1}{N-2}}\stackrel{\_}{\Phi }\left(\tilde{X}\right)\right]={\left[u_{\Phi }\tilde{P}\right]\left[\begin{array}{cc}{h}_{\Phi }& 0^T\\ \tilde{\Lambda }{\tilde{P}}^T\sqrt{\frac{N-1}{N\left(N-2\right)}}\left(\Phi \left(x_1\right)-{\tilde{m}}_{\Phi }\right)& \widetilde{\sum _{\Phi }^{}}\end{array}\right]\left[\begin{array}{cc}1& 0^T\\ 0_{N-1}& {\tilde{D}}_{\Phi }\end{array}\right]}^T={\left[u_{\Phi }\tilde{P}\right]\left[\begin{array}{cc}{h}_{\Phi }& 0^T\\ \tilde{\Lambda }{R}_{\Phi }^T{P}^T\sqrt{\frac{N-1}{N\left(N-2\right)}}\left(\Phi \left(x_1\right)-{\tilde{m}}_{\Phi }\right)& \widetilde{\sum _{\Phi }^{}}\end{array}\right]\left[\begin{array}{cc}1& 0^T\\ 0_{N-1}& {\tilde{D}}_{\Phi }\end{array}\right]}^T& \left(5\right)\end{array}$ where the regulating factor of the initial monitoring model in the continuous annealing process:

$\begin{array}{cc}\begin{array}{c}{h}_{\Phi }=\left(I-\tilde{P}\tilde{\Lambda }{\tilde{P}}^T\right)\sqrt{\frac{N-1}{N\left(N-2\right)}}\left(\Phi \left(x_1\right)-{\tilde{m}}_{\Phi }\right)\\ =\sqrt{\frac{N-1}{N\left(N-2\right)}}\left(I-\tilde{P}\tilde{\Lambda }{\tilde{P}}^T\right)\left(\Phi \left(x_1\right)-{\tilde{m}}_{\Phi }\right)\end{array}& \left(6\right)\\ \begin{array}{c}=\sqrt{\frac{N-1}{N\left(N-2\right)}}\Phi \left(x_1\right)-\frac{1}{N-1}\Phi \left(\tilde{X}\right)1_{N-1}-\\ \Phi \left(\tilde{X}\right)\tilde{A}\tilde{\Lambda }{\tilde{A}}^T\left({\Phi \left(\tilde{X}\right)}^T\Phi \left(x_1\right)-\frac{1}{N-1}{\Phi \left(\tilde{X}\right)}^T\Phi \left(\tilde{X}\right)1_{N-1}\right)\\ =\frac{N-1}{N\left(N-2\right)}\Phi \left(x_1\right)-\frac{1}{N-1}\Phi \left(\tilde{X}\right)1_{N-1}-\end{array}& \\ \begin{array}{c}\Phi \left(\tilde{X}\right)\tilde{A}\tilde{\Lambda }{\tilde{A}}^T\left(k\left(\tilde{X},x_1\right)-\frac{1}{N-1}K\left(\tilde{X}\right)1_{N-1}\right)\\ \frac{N-1}{N\left(N-2\right)}\Phi \left(x_1\right)-\Phi \left(\tilde{X}\right)B\\ =\sqrt{\frac{N-1}{N\left(N-2\right)}}\sqrt{1-2B^{T}k\left(\tilde{X},x_1\right)+B^{T}K\left(\tilde{X}\right)B}\end{array}& \\ \begin{array}{c}{u}_{\Phi }=\frac{1}{h_{\Phi }}\sqrt{\frac{N-1}{N\left(N-2\right)}}\left(I-\tilde{P}\tilde{\Lambda }{\tilde{P}}^T\right)\left(\Phi \left(x_1\right)-{\tilde{m}}_{\Phi }\right)\\ =\frac{1}{h_{\Phi }}\sqrt{\frac{N-1}{N\left(N-2\right)}}\left[\Phi \left(x_1\right)-\Phi \left(\tilde{X}\right)B\right]\end{array}& \left(7\right)\end{array}$ The correcting matrix of the main factors for the initial model in the continuous annealing process:

$\begin{array}{cc}B=\frac{1}{N-1}{1}_{N-1}+\tilde{A}\tilde{\Lambda }{\tilde{A}}^T\left(k\left(\tilde{X},x_1\right)-\frac{1}{N-1}K\left(\tilde{X}\right)1_{N-1}\right)& \left(8\right)\end{array}$ where K({tilde over (X)}) indicates the inner product of the middle matrix in the continuous annealing process, k({tilde over (X)}, x₁) indicates the inner product of {tilde over (X)} and x₁;Set

$\begin{array}{c}{V}_{\Phi }=\left[\begin{array}{cc}{h}_{\Phi }& 0^T\\ \tilde{\Lambda }{R}_{\Phi }^T{P}^T\sqrt{\frac{N-1}{N\left(N-2\right)}}\left(\Phi \left(x_1\right)-{\tilde{m}}_{\Phi }\right)& \widetilde{\sum _{\Phi }^{}}\end{array}\right]\\ =\left[\begin{array}{cc}{h}_{\Phi }& 0^T\\ \tilde{\Lambda }{R}_{\Phi }^T{P}^T\sqrt{\frac{N-1}{N\left(N-2\right)}}\left(\begin{array}{c}k\left(X,x_1\right)-\\ \frac{1}{N-1}K\left(X,\tilde{X}\right)1_{N-1}\end{array}\right)& \widetilde{\sum _{\Phi }^{}}\end{array}\right]\end{array}$ We get V_Φ=U_Φ′Σ_Φ′D′_Φ^T by singular value decomposition of V_Φ. U_Φ′ is the eigenvectors matrix, Σ_Φ′ is the diagonal matrix, D_Φ′ is the corresponding right-singular matrix. Substituting V_Φ into Equation (2) and we have:

$\begin{array}{cc}[\sqrt{\frac{N-1}{N\left(N-2\right)}}\left(\Phi \left(x_1\right)-{\tilde{m}}_{\Phi }\right)\sqrt{\frac{1}{N-2}}\hspace{1em}\stackrel{\_}{\Phi }\left(\tilde{X}\right)]=\left[\frac{1}{h_{\Phi }}\sqrt{\frac{N-1}{N\left(N-2\right)}}\left[\Phi \left(x_1\right)-\Phi \left(\tilde{X}\right)B\right]\Phi \left(\tilde{X}\right)\tilde{A}\right]\times U_{\Phi }^{\prime }\sum _{\Phi }^{\prime }{D_{\Phi }^{\mathrm{\prime T}}\left[\begin{array}{cc}1& 0^T\\ 0_{N-1}& {\tilde{D}}_{\Phi }\end{array}\right]}^T=\Phi \left(\left[x_1\tilde{X}\right]\right)\left[\begin{array}{cc}\frac{1}{h_{\Phi }}\sqrt{\frac{N-1}{N\left(N-2\right)}}& 0^T\\ \frac{1}{h_{\Phi }}\sqrt{\frac{N-1}{N\left(N-2\right)}}B& \tilde{A}\end{array}\right]\times U_{\Phi }^{\prime }\sum _{\Phi }^{\prime }{D_{\Phi }^{\mathrm{\prime T}}\left[\begin{array}{cc}1& 0^T\\ 0_{N-1}& {\tilde{D}}_{\Phi }\end{array}\right]}^T=\Phi \left(\left[\tilde{X}{x}_1\right)\right]\left[\begin{array}{cc}\tilde{A}& -\frac{1}{h_{\Phi }}\sqrt{\frac{N-1}{N\left(N-2\right)}}B\\ 0^T& \frac{1}{h_{\Phi }}\sqrt{\frac{N-1}{N\left(N-2\right)}}\end{array}\right]\times U_{\Phi }^{\prime }\sum _{\Phi }^{\prime }{D_{\Phi }^{\mathrm{\prime T}}\left[\begin{array}{cc}1& 0^T\\ 0_{N-1}& {\tilde{D}}_{\Phi }\end{array}\right]}^T& \left(9\right)\end{array}$ The main factors P of C^F can be expressed as

$\begin{array}{cc}\begin{array}{c}P=\Phi \left(\left[x_1\tilde{X}\right]\right)\left[\begin{array}{cc}\frac{1}{h_{\Phi }}\sqrt{\frac{N-1}{N\left(N-2\right)}}& 0^T\\ -\frac{1}{h_{\Phi }}\sqrt{\frac{N-1}{N\left(N-2\right)}}B& \tilde{A}\end{array}\right]U_{\Phi }^{\prime }\\ =\Phi \left(X\right)\left[\begin{array}{cc}\frac{1}{h_{\Phi }}\sqrt{\frac{N-1}{N\left(N-2\right)}}& 0^T\\ -\frac{1}{h_{\Phi }}\sqrt{\frac{N-1}{N\left(N-2\right)}}B& \tilde{A}\end{array}\right]U_{\Phi }^{\prime }\end{array}& \left(10\right)\end{array}$ And P=Φ(X)A, so we get Equation (11)

$\begin{array}{cc}A=\left[\begin{array}{cc}\frac{1}{h_{\Phi }}\sqrt{\frac{N-1}{N\left(N-2\right)}}& 0^T\\ -\frac{1}{h_{\Phi }}\sqrt{\frac{N-1}{N\left(N-2\right)}}B& \tilde{A}\end{array}\right]U_{\Phi }^{\prime }& \left(11\right)\end{array}$ From Equation (11), Ã can be calculated:

$\left[\begin{array}{cc}\frac{1}{h_{\Phi }}\sqrt{\frac{N-1}{N\left(N-2\right)}}& 0^T\\ -\frac{1}{h_{\Phi }}\sqrt{\frac{N-1}{N\left(N-2\right)}}B& \tilde{A}\end{array}\right]={A\left(U_{\Phi }^{\prime }\right)}^{-1}$ After the main factors P is obtained from the initial monitoring model of the continuous annealing process in Step 2 and we can get the score vector tεR^r in the feature space of the continuous annealing process.

$\begin{array}{cc}\begin{array}{c}t=P^T\left[\Phi \left(x_{\mathrm{new}}\right)-m_{\Phi }\right]\\ =A^T{\Phi \left(X\right)}^T{\Phi \left(X\right)}^T\left[\Phi \left(x_{\mathrm{new}}\right)-\frac{1}{N}\Phi \left(X\right)1_N\right]\\ =A^T\left[k\left(X,x_{\mathrm{new}}\right)-\frac{1}{N}K\left(X\right)1_N\right]\end{array}& \left(12\right)\end{array}$ where P=[p₁, p₂, . . . . , p_r], r is the number of the retaining nonlinear principal component, k(X, x_new) indicates the inner product of the sample matrix X and the new sample x_new in the continuous annealing process. T² and SPE statistics of the new samples x_new are calculated by Equation (13) and (14).T₁²=t^TΛ⁻¹t  (13)SPE₁=[Φ(x_new)−m_Φ]^T(I−PP^T)[Φ(x_new)−m_Φ]  (14)where Λ is the eigenvalues matrix of the principal component. T² statistic satisfies the F distribution:

$T^2=\frac{r\left(N^2-1\right)}{N\left(N-r\right)}{F}_{r,N-r}$ Among them, N is the number of the sample, r is the number of the retaining principal component, the upper limit of the T² statistic is

$\begin{array}{cc}{T}_{\beta }^2=\frac{r\left(N^2-1\right)}{N\left(N-r\right)}{F}_{r,N-r,\beta }& \left(15\right)\end{array}$ Among them, β is the confidence level, while the Q statistic meets the χ² distribution, the control upper limit isQ_β=gχ²(h)  (16)Among them, g=ρ²/2μ,h=2μ²/ρ², μ and ρ² indicate the sample mean and variance corresponding Q statistic. If T₁² and SPE₁ are greater than their respective confidence, an alarm occurs, which indicates the continuous annealing process anomalies occur. Otherwise go to step 3.

Update the initial monitoring model of the continuous annealing process of step 2 using the recursive kernel principal component analysis stated by step 3, and calculate the main factors {circumflex over (P)} after updating the continuous annealing process model. The method is as follows:

x_new is a new samples in the continuous annealing process and can be used, Φ(x_new) is the new samples x_new's projection in the feature space in the continuous annealing process, Φ(X_new)=Φ([{tilde over (X)} x_new]) is the samples matrix's projection in the feature space in the updated continuous annealing process, the mean matrix {circumflex over (m)}_Φ of Φ(X_new) and covariance matrix Ĉ^F are respectively

$\begin{array}{cc}{\hat{m}}_{\Phi }=\frac{1}{N}\Phi \left(\left[\tilde{X}{x}_{\mathrm{new}}\right]\right)1_N=\frac{N-1}{N}{\tilde{m}}_{\Phi }+\frac{1}{N}\Phi \left(x_{\mathrm{new}}\right)& \left(17\right)\\ \begin{array}{c}{\hat{C}}^F=\frac{1}{N-1}\stackrel{\_}{\Phi }\left(\left[\tilde{X}{x}_{\mathrm{new}}\right]\right){\stackrel{\_}{\Phi }\left(\left[\tilde{X}{x}_{\mathrm{new}}\right]\right)}^T\\ =\frac{N-2}{N-1}\left[\sqrt{\frac{N-1}{N\left(N-2\right)}}\left(\Phi \left(x_{\mathrm{new}}\right)-{\tilde{m}}_{\Phi }\right)\sqrt{\frac{1}{N-2}}\stackrel{\_}{\Phi }\left(\tilde{X}\right)\right]\times \\ {\left[\sqrt{\frac{N-1}{N\left(N-2\right)}}\left(\Phi \left(x_{\mathrm{new}}\right)-{\tilde{m}}_{\Phi }\right)\sqrt{\frac{1}{N-2}}\stackrel{\_}{\Phi }\left(\tilde{X}\right)\right]}^T\end{array}& \left(18\right)\end{array}$ From the equation (2) to (9) we can get

$V_{\Phi }^{\prime }=\left[\begin{array}{cc}\widetilde{\sum _{\Phi }^{}}& \tilde{\Lambda }{\tilde{A}}^T\sqrt{\frac{N-1}{N\left(N-2\right)}}\left(k\left(\tilde{X},x_{\mathrm{new}}\right)-\frac{1}{N-1}K\left(\tilde{X}\right)1_{N-1}\right)\\ 0^T& h_{\Phi }^{\prime }\end{array}\right]$ We can get V_Φ′=U_ΦⁿΣ_ΦⁿD_Φ^nT by singular value decomposition of v_Φ′And thus we can get the main factors {circumflex over (P)} and the engenvalues matrix {circumflex over (Λ)} of Ĉ^F

$\begin{array}{cc}\hat{P}=\Phi \left(\left[\begin{array}{cc}\tilde{X}& x_{\mathrm{new}}\end{array}\right]\right)\left[\begin{array}{cc}\tilde{A}& -\frac{1}{h_{\Phi }^{\prime }}\sqrt{\frac{N-1}{N\left(N-2\right)}}{B}^{\prime }\\ 0^T& \frac{1}{h_{\Phi }^{\prime }}\sqrt{\frac{N-1}{N\left(N-2\right)}}\end{array}\right]U_{\Phi }^{″}=\Phi \left(X_{\mathrm{new}}\right)\hat{A}& \left(19\right)\\ \hat{\Lambda }=\frac{N-2}{N-1}{\Sigma }_{\Phi }^{″2}& \left(20\right)\end{array}$ where the regulating factor of the main factors for the updating monitoring model in the continuous annealing process:

$\begin{array}{cc}{h}_{\Phi }^{\prime }=\frac{N-1}{N\left(N-2\right)}\sqrt{1-2B^{\prime T}k\left(\tilde{X},x_{\mathrm{new}}\right)+B^{\prime T}K\left(\tilde{X}\right)B^{\prime }}& \left(21\right)\end{array}$ The correcting matrix of the main factors for the updating monitoring model in the continuous annealing process:

$\begin{array}{cc}{B}^{\prime }=\frac{1}{N-1}{1}_{N-1}+\tilde{A}\tilde{\Lambda }{\tilde{A}}^T\left(k\left(\tilde{X},x_{\mathrm{new}}\right)-\frac{1}{N-1}K\left(\tilde{X}\right)1_{N-1}\right)& \left(22\right)\end{array}$ k( X, x_new) indicates the inner product of the middle matrix X and the new sample x_new;

By using Hotelling's T² statistic and squared prediction error (SPE) statistic for fault monitoring stated in step 4, the determining methods of T² and squared prediction error (SPE) statistics are as follows:

For a new sample z in the continuous annealing process, its score vector tεR^r in the feature space is

$\begin{array}{cc}\begin{array}{c}t={\hat{P}}^T\left[\Phi \left(z\right)-{\hat{m}}_{\Phi }\right]\\ ={\hat{A}}^T{\Phi \left(X_{\mathrm{new}}\right)}^T\left[\Phi \left(z\right)-\frac{1}{N}\Phi \left(X_{\mathrm{new}}\right)1_N\right]\\ ={\hat{A}}^T\left[k\left(X_{\mathrm{new}},z\right)-\frac{1}{N}K\left(X_{\mathrm{new}}\right)1_N\right]\end{array}& \left(23\right)\end{array}$ Among them, {circumflex over (P)}=[{circumflex over (p)}₁, {circumflex over (p)}₂, . . . , {circumflex over (p)}_r], r is the number of the retaining principal component, k(X_new, z) indicates the inner product vector of the updating samples matrix X_new and the new samples X_new in the continuous annealing process. T₂² and SPE₂ statistics of the new sample z in the continuous annealing process are calculated from the equation (24) and (25).T₂²=t^T{circumflex over (Λ)}⁻¹t  (24)SPE₂=[Φ(z)−{circumflex over (m)}_Φ]^T(I−{circumflex over (P)}{circumflex over (P)}^T)[Φ(z)−{circumflex over (m)}_Φ]  (25)where {circumflex over (Λ)} is the variance matrix of the principal component;The confidence limits of T₂² and SPE₂ statistics of the new sample z can be obtained by the equation (15) and (16). If T₂² or SPE₂ statistics are greater than their confidence limits, we think there is a fault and an alarm will occur. Otherwise go to step 3;

Advantages of the invention: the invention proposes a fault detection method of the continuous annealing process based on the recursive kernel principal component analysis mainly to solve the nonlinear and time-varying data problem. RKPCA updates the model by recursively computing the eigenvalues and main factors of training data covariance. The process monitoring results by using the method shows that the method can not only greatly reduce false alarms, but also improve the accuracy of the fault detection.

BRIEF DESCRIPTION OF THE DRAWINGS

FIG. 1 shows a Physical layout of the continuous annealing process of a fault detection method based on the recursive kernel principal component analysis according to the present invention;

FIG. 2 shows a total flow figure of the fault detection method in the continuous annealing process based on the recursive kernel principal component analysis in the invention;

FIG. 3 shows a model flow picture in the continuous annealing process based on the recursive kernel principal component analysis in the invention;

FIG. 4 shows the T² statistic of the fault detection method in the continuous annealing process based on the recursive kernel principal component analysis in the invention;

FIG. 5 shows the SPE statistic of the fault detection method in the continuous annealing process based on the recursive kernel principal component analysis in the invention;

FIG. 6 shows the number of the calculated principal component in the fault detection method in the continuous annealing process based on the recursive kernel principal component analysis in the invention;

FIG. 7 shows the T² statistic of the fault detection method in the continuous annealing process based on the recursive kernel principal component analysis in the invention; and

FIG. 8 shows the SPE statistic of the fault detection method in the continuous annealing process based on the recursive kernel principal component analysis in the invention.

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENT

We illustrate further in detail combining of the following drawings and examples for the invention.

The physical layout of the continuous annealing process is shown in FIG. 1. Where the width of the strip is 900-1230 mm, the thickness is 0.18-0.55 mm, the maximum line speed is 880 m/min, the maximum weight is 26.5 t and is heated to 710° C. The first entry coil is opened by payoff reel (POR), and it is welded into a strip finally. The strip passes through 1# bridle roll (1BR), entry loop (ELP), 2# bridle roll (2BR), 1# dancer roll (1DCR), and 3# bridle roll (3BR), then it enters the continuous annealing furnace. The annealing technologies consist of: rapid cooling-reheating-inclined over ageing. The annealing equipments include heating furnace (HF), soaking furnace (SF), slow cooling furnace (SCF), 1# cooling furnace (1C), reheating furnace (RF), over ageing furnace (OA), and 2# cooling furnace (2C). After completing the annealing craft, the strip in turn pass through 4# bridle roll (4BR), delivery loop (DLP), and 5# bridle roll (5BR), and temper rolling machine (TPM), 6# bridle roll (6BR), 2# dancer roll (2DCR), and 7# bridle roll (7BR). Finally, the strip enters roll type reel (TR) to become coil.

The invention is the fault detection method of the continuous annealing process based on the recursive kernel principal component analysis, shown in FIG. 2, including the following steps:

Step 1: Collect data and standardize the collecting data, collecting data in the continuous annealing industrial process including: the roll speed, current and tension of entry loop (ELP), which includes 37 roll speed variables, 37 current variables, and 2 tension variable in both sides of ELP;

There are a total of 76 process variables of ELP in the continuous annealing process. There are 200 history samples. Also, there are 300 real time samples. 99% confidence limits are selected. Each sample contains 76 variables. Some sample data are shown in Table 1 and Table 2, and ten sets of data randomly selected from the training data and test data is shown in Table 1 and Table 2.

TABLE 1
Ten sets of the history data of ELP
	Var.
		1R Roller		5R Roller		18R Roller	TM1	TM2
No.	1R Current	Speed	5R Current	Speed	18R Current	Speed	Tension	Tension
1	0.658610	36.4698	38.2673	747.034	31.0010	657.585	6.73518	6.70954
2	0.663005	36.6896	37.6600	746.140	31.4893	658.350	6.66926	6.85237
3	0.670451	36.5461	35.5238	746.864	30.0122	657.606	6.74617	6.69856
4	0.670451	36.1189	39.4514	746.736	30.3754	658.159	7.10506	6.92927
5	0.663493	36.2135	36.8147	746.715	29.6674	657.861	6.65827	6.51179
6	0.664347	36.4179	36.1402	747.034	30.7416	657.904	6.43854	6.73518
7	0.664958	36.2196	38.4047	746.396	31.1139	657.734	6.52277	6.39826
8	0.659953	36.1097	37.1046	746.587	30.7691	658.159	6.45319	6.62898
9	0.662028	36.4332	36.3661	746.162	30.5676	658.201	6.56306	6.39826
10	0.664225	36.1921	37.9866	746.672	30.7416	657.670	6.88167	6.53376

TABLE 2
Ten sets of the real time data of ELP a
	Var.
		1R Roller		5R Roller		18R Roller	TM1	TM2
No.	1R Current	Speed	5R Current	Speed	18R Current	Speed	Tension	Tension
1	0.656535	35.7221	37.8920	745.694	29.9603	657.797	6.37263	6.81941
2	0.624918	35.4292	37.9347	736.379	25.1995	650.845	6.69490	6.61433
3	0.631876	34.9012	41.1604	726.468	27.5799	640.448	6.48249	7.02815
4	0.683879	34.5380	46.0769	720.811	31.4801	631.986	6.30304	6.79377
5	0.649088	34.2237	46.5621	718.833	34.0131	631.178	6.24811	7.06111
6	0.686442	34.2542	50.6851	715.600	35.4231	626.947	6.32135	6.72786
7	0.688273	34.0864	51.7044	716.174	35.5421	627.351	6.39094	6.86336
8	0.687419	34.1535	52.4125	715.919	33.1891	627.372	6.66926	6.72053
9	0.690959	34.2634	53.9841	715.898	33.0396	627.159	6.44587	6.63264
10	0.700481	34.2359	55.3880	715.515	32.2095	627.308	6.50080	6.77546

Step 2: Build the initial monitoring model of ELP in the continuous annealing process and calculate the main factor P of fault in the continuous annealing process and determine confidence limits by using 200 samples after standardized samples in Step 1. Monitor a new sample x_new of continuous annealing process. If it is abnormal, an alarm will be given, otherwise go to Step 3.

Set 200 samples of the continuous annealing process as the matrix X, and the latter 199 data of the samples as themediate matrix {tilde over (X)}. They are mapped to high dimensional feature space by the projection Φ; Find the transmission factor à of the middle matrix, according to equations (2) and (10), we can get the covariance matrix C^F and the main factor P of the sample matrix X by calculating and the T₁² and SPE, statistics of the new sample x_new in continuous annealing process using the main factors P according to equations (13) and (14) and determine whether they are greater their respective confidence limit. There is no fault by calculating and go to step 3.

Step 3: Use recursive kernel principal component analysis method to update the initial monitoring model of continuous annealing process in Step 2 and calculate the main factor {circumflex over (P)} of fault in the continuous annealing process after updating continuous annealing model according to Equation (19);

x_new is a new sample of the continuous annealing process, Φ(x_new) is the mapping to the feature space of the new sample x_new of the continuous annealing process. Φ(X_new)=Φ([{tilde over (X)} x_new]) is the updating sample matrix of the continuous annealing process and the transmission factor  and eigenvalues matrix  of the updated covariance Ĉ^F can be calculated by equations (19) and (20), respectively. Thus we can get the main factor {circumflex over (P)} of the updating sample matrix in the continuous annealing process. Here we randomly select ten sets of data of the transmission factor, as shown in Table 3.

TABLE 3
Ten sets of data of the transmission factor Â
−0.002234	0.116932	0.02946	0.004251	−0.031282	0.00692	−0.14756	−0.000921
0.0202367	−0.00366	0.00288	−0.00933	0.0136258	0.06262	0.073793	−0.040450
0.0326804	−0.05354	−0.0026	0.009399	0.0695079	0.09472	0.082283	−0.057724
−0.025156	−0.00394	0.06529	0.056101	0.0082486	−0.0995	−0.16148	−0.014453
0.0546749	−0.01765	−0.0377	−0.04835	0.0261276	0.05037	0.033161	0.0056228
0.0020289	0.04197	0.01145	−0.02075	−0.075466	−0.1277	−0.15009	0.0565441
0.0031576	−0.12491	−0.0031	0.029215	0.1613627	0.09810	0.067874	−0.1144526
0.0374927	−0.01329	0.03562	−0.00175	−0.031035	−0.09767	−0.10696	0.0197142
0.0436574	−0.07263	0.01474	−1.80547	0.098647	0.038714	−0.01312	−0.0633847
0.0876383	0.048630	−0.01972	−0.10252	−0.01356	0.028529	−0.06496	−0.0162924

Step 4: Fault Detection by Using the Updated Continuous Annealing Process Model;

By using Hotelling's T² statistic and squared prediction error (SPE) statistic for fault detection, we can determine whether the fault of the continuous annealing process occurs. When the T² statistic or SPE statistic are beyond their respective confidence limit, we think that there is a fault; otherwise the whole process is normal and goes to Step 3 and continues to update the monitoring model.

RKPCA uses 200 history samples of the continuous annealing process to build the initial model and then we can update the model according to 300 real time data. We use the T² and SPE statistics in order to monitor the process. For a new sample z among 300 real time data, its score vector t in the feature space can be computed by Equation (23). The T² and SPE statistics of the new sample z are calculated by Equation (24) and (25), and then we can determine their confidence limits according to Equation (15) and (16). When the T² statistic and SPE statistic are greater than their confidence limit we think that there is failure and an alarm is given. On the contrary, the whole process is normal. Go to step 3 and continue to update the monitoring model. The monitoring results for the continuous annealing process by calculating are shown in FIG. 4 and FIG. 5. We can see that the T² and SPE statistics generated by RKPCA method exceed the confidence limit at sample 175. In fact, the beak strip fault is introduced at sample 175. The simulation results shows that the proposed RKPCA method by updating recursively the model ensure that the model's effectiveness in the process of change so that we can monitor timely the continuous annealing process failure. FIG. 4 and FIG. 5 shows that the confidence limit based on the RKPCA method has also been updated. FIG. 6 shows the changes of the number of the retained principal component. On the contrast, when KPCA method is used to monitor the continuous annealing process, the model can't be updated recursively. The generated T² and SPE statistics are shown in FIGS. 7 and 8. In the first stage, due to the continuous annealing process's instability, the T² and SPE statistics exceed temporarily their respective confidence limits, but in the stable process, there is no fault.

The above simulation example shows in the invention—the effectiveness of the fault detection in the continuous annealing process based on the recursive kernel principal component analysis and realizes monitoring of the continuous annealing process.

Claims

1. A fault detection method in a continuous annealing process based on a recursive kernel principal component analysis (RKPCA), comprising the following steps: Step 1: collecting data and standardizing samples using a processor by detecting roll speed, current and tension of an entry loop (ELP);Step 2: extracting principal factors P of the fault in the continuous annealing process using the processor by building an initial monitoring model of the continuous annealing process with N standardized samples in Step 1; monitoring a new sample xnew of the continuous annealing process, and if it is abnormal, generating an alarm, otherwise going to Step 3; wherein the extracted principal factor P in the continuous annealing process is as follows:

2. The method as claimed in claim 1, wherein the Step 2 of building the initial monitoring model of the continuous annealing process in the Step 1 includes the following steps: updating, using the processor, recursively eigenvalues in the feature space of the sample covariance matrix by the RKPCA;letting, using the processor, X=[x1, x2, . . . , xN] be the sample matrix of the continuous annealing process, wherein x1, x2, . . . , xN are the samples of the continuous annealing process, N is the sample number, {tilde over (X)}=[x2, . . . , xN]εRm×(N-1) is the middle matrix of the continuous annealing process, m is the number of sampling variables in the continuous annealing process, Xnew[{tilde over (X)} xnew] is the sample matrix of updating model in the continuous annealing process, xnew is the new sample of the continuous annealing process; mapping X, {tilde over (X)} and Xnew to the high-dimensional feature space to become Φ(X), Φ({tilde over (X)}) and Φ(Xnew), respectively, so the mean vector mΦ and covariance matrix CF of Φ(X) can be calculated

3. The method as claimed in claim 1, wherein the Step 3 of updating the initial monitoring model of the continuous annealing process built in the Step 2 and calculating the principal factor {circumflex over (P)} of the fault in the updated continuous annealing process model by the RKPCA includes: letting, using the processor, xnew be new samples in the continuous annealing process and be capable of being used, Φ(xnew) be the new samples xnew's projection in the feature space in the continuous annealing process, Φ(Xnew)=Φ([{tilde over (X)} xnew]) be the samples matrix's projection in the feature space in the updated continuous annealing process, the mean matrix {circumflex over (m)}Φ of Φ(Xnew) and covariance matrix CF are given by respectively