Research On Process Monitoring Method Based On Kernel Principle Component Analysis

The nonlinear characteristics exist between the variables of samples collected in industrial processes. Using the linear method can result in extraction of the data is not very adequate. There are often multiple operating modes in the nonlinear industrial processes, and the process data in different operating modes show different statistical properties. The global modeling method does not analysis the special characteristics of each mode. When a fault is detected later, we hope to find the cause of the failure, and diagnose fault. Traditional KPCA-based fault reconstruction method only extracts fault direction by analyzing fault data. It ignores the relationships between the fault data and the normal data. In response to these problems, this paper mainly does the following research work:(1) This paper presents multimode fault detection method based on local kernel principal component analysis algorithm (LKPCA). This method cluster data into different modes by k-means clustering and obtain the weight matrix W reflected its mode category. The data is mapped into a high dimensional space. W and kernel method are used to keep the covariance of each mode data biggest. That can extract the common characteristic reflecting public information of the data. Then, special characteristic was classified into different modes. Monitor the public information and the specific information respectively. LKPCA based multimode fault detection method is used in the production process of fused magnesia to verify that the method can effectively solve the multimode process fault detection. Compared with the KPCA and local PCA methods also prove the superiority of the method.(2) This paper presents a fault reconstruction method based on Fault-relevant KPCA (FRKPCA). This method saves normal data and fault data as historical data. The data is nonlinearly mapped to high-dimensional feature space. The fault data is decomposed into subspace (PCS) and residual subspace (RS) by KPCA model of normal data. By analyzing the failure information on the impact of normal data to extract fault direction leading to T2and SPE overrun control limits in each subspace. Fault data is corrected using reconstruction model to eliminate the phenomenon of statistics overrun. Use fault reconstruction proposed for fault diagnosis, and use it for online fault diagnosis of fused magnesium process. It can prove the validity of the reconstruction method.