A topology model based on common and specific feature separation for multimode process monitoring

Abstract

Process production often operates under multiple modes due to factors such as market demands and environmental policies. However, since varying mode operations are based on the same production equipment, they share a common structural correlation. Conversely, due to the varying setpoints and production goals among modes, there are diverse correlations between variables among modes. This paper proposes a topology model based on common and specific feature separation (TMCSFS) for multimode process monitoring. In order to achieve an accurate description of process operating state, this model uses the shared common correlation given by P&ID and expert knowledge, and the specific correlation learned by a self-attention mechanism to extract deep topological features. And these two correlation graphs will also be used in online mode identification. Experiments on Tennessee Eastman process (TE process) showed that features extracted by TMCSFS could fully characterize the topological changes in process operation, reflect the recovery capability of process feedback loop, and enhance the explainability of TMCSFS.