A Fine-scale Industrial Process Distribution Evaluation Method with Concurrent Static and Dynamic Dissimilarity Analysis

Abstract

Dissimilarity (DISSIM) analysis is a widely used method that can detect the changes of data distribution at the early stage, whereas conventional distance-based monitored statistics may still keep inside of the normal region. However, temporal distribution should be isolated from steady-state information, which are representations of process dynamics and provides precise discrimination of normal deviations from real faults. In this study, a concurrent static and dynamic dissimilarity analysis based on slow feature analysis technique (CSDDISSIM) is developed for identifying incipient faults sensitively and exploring physical interpretation. First, static features and their temporal counterpart are obtained for distribution exploration. Then the changes of distributions are checked in fine scale and the monitoring scheme is developed to distinguish the current process status, which are normal, operating deviation, dynamic abnormality and real fault. In this way, the industrial process status can be captured with beneficial process interpretation. The validity of the proposed method is tested in a typical benchmark, namely Tennessee Eastman process.