Novel data segmentation methods for data driven process analyses

Abstract

With the advent of Computer based Control Systems, e.g., Distributed Control System (DCS), process industries are storing huge amount of data everyday. For example, if data is collected at a sampling rate of 4s, there will be 21,600 samples for each variable every day. These data are a rich source of information. Often these routine operating data are used for calculating or estimating various performance metrics of the control loops or the process. Sometimes, conventional approach of computing a performance metric using all available data points may lead to erroneous conclusion due to the presence of various unexpected events, such as abnormal pulses, temporary sensor failures, transmitter failures, non-stationary trends or process disturbances. In this study, novel data segmentation methods have been developed to increase the reliability of data driven process analyses. Calculation of a performance measure or metric requires only a few hundred or a couple of thousand data points. Therefore, available data can be divided into several windows for metric estimation. A three steps novel data segmentation method has been developed. In the first step, a required data window length is determined for estimating a performance metric. In the second step, the time series is divided into several data windows for metric estimation. Finally, an index named reliability index (RI) is calculated from the results obtained in different data windows. Reliability index (RI), bounded between 0 and 1, calculates the probability of getting similar results in different data windows. A higher value of reliability index (RI) indicates more reliable estimate of the metric. The proposed novel data segmentation method has been used successfully for oscillation detection, nonlinearity analysis and performance index calculation from routine operating data.