Extraction of Information from Time-Series Data

点击翻译按钮获取中文摘要

Abstract - Joseph and Motard - 9529578 Many chemical plants have distributed computer control systems (DCS) which collect time series data from thousands of sensors. The data contained in this past operational history is used to monitor, control and optimize the plant operation. The objective of this work is to develop and test tools and algorithms for extracting information from time series data generated by on-line process sensors. Information extraction tasks include converting raw sensor data to meaningful process variables, identifying process models such as input-output transfer functions, monitoring the status of plant operation and detection of abnormal events. The PIOs will investigate the problem at the sensor level, the unit level and the plant level. Specific issues addressed by this project include: (I) the problem of sensor fingerprinting, i.e. the problem of extracting information about the underlying process phenomena from the raw signal generated by the sensor, (ii) the problem of improving the accuracy of parameters in process models generated from noisy input/output time series data by selective prefiltering to maximize the signal to noise ratio, and (iii) the problem of event detection and generation of an event data base containing information needed to monitor and diagnose performance of the plant from various points of view including quality, control, economics, environmental, maintenance, and fault detection.