Feature Extraction from Time-Series Data for Process Monitoring

A plant operator monitors time-series data of process variables to judge the state of process, diagnose abnormal states, and to identify failure origins. In this study, a new feature extraction method which extracts simultaneously both local features such as spikes and step changes, and the trend which characterizes global changes is provided from the viewpoint of process monitoring. In this method, the continuous function interpolated from the time-series data is represented by a series of inflection points first. Each time interval between two inflection points is called an episode. Then an approximation function of the time-series data is made iteratively by way of merging these episodes. This feature extraction method is also useful for compaction of a large number of process data.