A Nonlinear Quality-relevant Process Monitoring Method with Kernel Input-output Canonical Variate Analysis

Abstract

Traditional process monitoring methods based on kernel canonical variate analysis do not extract variances. They cannot judge whether a process fault that is detected affects product quality. A nonlinear quality-relevant process monitoring method based on kernel input-output canonical variate analysis (KIOCVA) is proposed. Firstly, Process variables and quality variables are mapped into higher-dimensional linear feature spaces via unknown nonlinear mappings respectively. The higher-dimensional linear feature spaces are projected to three subspaces, an input-output correlated subspace that captures correlations between process data and quality data, an uncorrelated input subspace and an uncorrelated output subspace. To monitoring the variances of the uncorrelated input subspace and the uncorrelated output subspace, principal component analysis is performed. Correlations and variances in the higher-dimensional linear feature spaces are extracted by means of nonlinear kernel functions. The proposed KIOCVA method can judge the process fault that is detected affects product quality or not. The effectiveness of the proposed method is demonstrated by case studies of Tennessee Eastman process.