(487b) High-Order Non-Stationary Information Extraction for Industrial Process Monitoring Using Multi-Cointegration

Process monitoring has gained considerable attention in previous decades as an effective method of preventing accidents and improving the stability and safety during real-time industrial process operation. With the widespread application of Distributed Control System (DCS), a large amount of production data can be stored, which provides abundant support for data-driven process monitoring methods. Multivariate statistical process monitoring (MSPM) as a common data-driven method, has been widely used in the monitoring of stationary process. However, some process variables show significant non-stationary characteristics due to process internal mechanism and external disturbances, for which MSPM may not be applicable. Although the non-stationary trends of these variables may be various, there are long-term equilibrium relationships among them, which can be extracted by cointegration analysis (CA). Therefore, CA has already been introduced to process monitoring by monitoring the cointegration relationship among non-stationary variables. However, in industrial practice, the process non-stationarity of variables is not only contributed by linear deterministic trends, but also certain high-order process information, such as the periodic characteristics. In this way, industrial process could be high-order non-stationary, and the cointegration relationship among variables from different orders is not fully considered by traditional CA, resulting in insufficient extraction of non-stationary information. To deal with this issue, a new non-stationary process monitoring strategy based on multi-cointegration is proposed in this work to broaden the application range of CA. The variables are first classified by the differential operation and stationary test according to different integrated orders. CA is then performed on them to extract the long-term equilibrium relationships and multi-cointegration relationships, based on which the monitoring statistics is constructed to determine whether the system is faulty. The proposed strategy is validated through a numerical case and an actual industrial case, and the results compared with other non-stationary process monitoring methods demonstrate its effectiveness in monitoring non-stationary process including high-order non-stationary variables.