Dual-step subspace partition for process monitoring

Available sensing measurements in modern industrial process include two significant characteristics: distribution and autocorrelation. Different types of sensing measurements exhibit different characteristics. Moreover, different feature extraction methods are suitable for data with corresponding characteristics. This paper proposes a novel dual-step subspace partition method in order to establish accurate monitoring modeling. In the first step, process variables are divided into Gaussian distribution subspace and non-Gaussian distribution subspace according to data distribution. The second step is conducted on two subspaces obtained in the first step. The Durbin-Watson testing method is used to test the strength of autocorrelation. As a consequence, four subspaces can be obtained: Gaussian and non-autocorrelation (GNA) subspace, Gaussian and autocorrelation (GA) subspace, non-Gaussian and non-autocorrelation (NGNA) subspace, non-Gaussian and autocorrelation (NGA) subspace. Proper feature extraction methods (such as, principal component analysis and independent component analysis) corresponding to distribution and autocorrelation are implemented to monitor each subspace. Then, the comprehensive monitoring indexes are constructed based on Bayesian inference and statistics in each subspace. Finally, the superiority and effectiveness of the proposed method is proved under the Tennessee Eastman process.