A Novel Dynamic Monitoring Method Based on Co-integration Analysis for Nonstationary Processes

Due to reasons such as misoperation and mode transition, most modern industrial processes exhibit nonstationary characteristics. Traditional multivariate statistical process monitoring (MSPM) methods like the principal component analysis (PCA) and partial least squares (PLSs) cannot capture essential information during the feature extraction processes under nonstationary conditions. To monitor nonstationary processes accurately, a novel co-integration analysis (CA)-based MSPM method for nonstationary process monitoring is proposed in this article. First, a stationary dataset is formed by combining the stationary variables within the original dataset and the equilibrium error sequences obtained by CA. Then, the dynamic variable PCA (DVPCA) is proposed for process monitoring based on the stationary dataset. The retained principal components (PCs) of the DVPCA are determined by their ability to distinguish between fault data and normal data. Two types of PCA models are established in the DVPCA, corresponding to $T^{2}$ statistic and Q statistic, respectively. Furthermore, the DVPCA captures the dynamic characteristics of processes by embedding the past information into the data matrix. A numerical case and two industrial experiments demonstrate the superiority of the proposed method.