Quality-relevant Process Monitoring Method Based On Canonical Variate Analysis

Process monitoring is of great significance to ensure that the industrial systems operate safely and efficiently.Multivariate statistical methods have been widely applied in the process monitoring field,but they mainly use process data for analysis.Process operation condition and product quality are related closely.A new quality-relevant process monitoring method combined with the canonical variate analysis(CVA)method is researched.The main work of this paper is as follows:Because large variance may exist in the residual space of traditional CVA model,the monitoring effect of squared prediction error(SPE)statistics is affected.And CVA-based method cannot monitor output data space which is unpredictable from the input data.An input-output CVA(IOCVA)process monitoring method is proposed in this paper.Process variables and quality variables are mapped into 5 subspaces: an input-output correlated subspace,an uncorrelated input principal subspace,an uncorrelated input residual subspace,an uncorrelated output principal subspace and an uncorrelated output residual subspace.Canonical variate analysis is utilized to extract the correlation between input and output data.The variance in the residual data space is also be considered.The method can judge whether the process fault impact product quality or not.Nonlinear relationships within variables widely exist in the actual industrial process.IOCVA method is extended to nonlinear cases by combining with the kernel trick.Kernel IOCVA(KIOCVA)method is proposed.First,the original data spaces are projected to high-dimensional feature spaces,and the canonical variates of high-dimensional spaces are extracted by using kernel CVA.Then,kernel functions are applied to deal with the reconstruction problem of high-dimensional residual spaces.Meanwhile,variance analysis is used to extract the principal components.Next,statistics are constructed in each subspace for a complete and precise fault detection.The simulation results in Tennessee Eastman(TE)process prove that the KIOCVA fault detection method is effective.