Process Monitoring Without Data Centralization: Least Squares Reconstruction

Centralization is indispensable to covariance-based multivariate statistical process monitoring (MSPM) methods, such as principal component analysis (PCA) and canonical correlation analysis (CCA). However, for most processes, the variable expectations rarely remain unchanged, and they may slowly change or step change because of dynamic and multimode features. Though partial least squares (PLS) avoids centralization by using least squares estimation (LSE), it can only identify the relationships between variables in different datasets. We propose a process monitoring method, least squares reconstruction (LSR), which can identify the relationships between variables in different datasets and in the same datasets simultaneously. In addition, LSR avoids centralization, and hence, it can be used for monitoring dynamic and multimode processes. The results of a simulated process and a gas fractionation process show that LSR can effectively handle the issue of varying variable expectations and is more sensitive to process faults.