Concurrent static and dynamic dissimilarity analytics for fine-scale evaluation of process data distributions

Abstract

Dissimilarity algorithm has been widely used for timely identifying changes of data distribution in fine scale while many distance-based monitored indexes still stay inside the normal region. However, it ignores the information of temporal distribution, thus fails to monitor operating conditions and process dynamics separately. In order to detect incipient faults sensitively and also provide more benefiting process comprehension, concurrent static and dynamic dissimilarity analytics based on slow feature analysis technique (CSDDISSIM) is developed in this work. The distributions of process status and dynamics are evaluated in fine scale. First, static features and their temporal counterpart are extracted, in which slow and fast varying information is separated for distribution evaluation. Then the changes of both static and dynamic distributions are checked and the monitoring policy is developed to distinguish different statues, including normal, static deviation, dynamic abnormality, and concurrent deviations. In this way, the industrial process status can be captured with a beneficial interpretation. The practical utility and efficacy of the proposed method are illustrated in the application to a real thermal power plant process.