Novel Quality-Relevant Process Monitoring based on Dynamic Locally Linear Embedding Concurrent Canonical Correlation Analysis

Quality-relevant process monitoring has emerged as a powerful tool in ensuring product quality for industrial processes. However, most industrial processes exhibit dynamic and nonlinear characteristics. A novel quality-relevant monitoring named dynamic locally linear embedding concurrent canonical correlation analysis (DLCCCA) is proposed for nonlinear dynamic processes. First, local structure information on the time-lagged process and quality variables is extracted through locally linear embedding (LLE). Second, the extracted structure information is incorporated into canonical correlation analysis (CCA) through formulating a new optimization objective. Consequently, the process data space is decomposed into predictable quality subspace, quality-irrelevant process subspace, and unpredictable quality subspace. Third, the unpredictable quality subspace and quality-irrelevant process subspace are further decomposed using principal component analysis (PCA). From the decomposed subspaces, Hotelling's T-2 and Q statistics are established for quality-relevant process monitoring. Moreover, kernel density estimation (KDE) is utilized to determine the corresponding threshold for each statistic. DLCCCA combines the merits of both LLE and CCA. Two case studies are carried out to illustrate the superior performance of DLCCCA by comparison with other relevant methods.