(375f) Time Delayed Cointegration Analysis for Non-Stationary Process Monitoring

During real-time industrial chemical operation, certain measurements show significant non-stationary trends due to process inherent mechanism, adjustment of operation conditions, external disturbances, and so on, which could disturb the extraction of process information through data-driven models and present a challenge to the determination of operating status. Cointegration analysis (CA) has been previously introduced to address this issue by extracting the long-term equilibrium relationships (LTER) among non-stationary series originated from process mechanism. However, during the continuous production process, there could exist time lags among variables, such as temperature and pressure, due to the varying locations of measuring instruments and the time taken for the fluid to flow. It will lead to the insufficient extraction of the cointegration relationship and limit the process monitoring performance. In this work, time delayed CA is proposed for non-stationary process monitoring. The variables are first divided into stationary and non-stationary parts, and time delayed mutual information (TDMI) is performed to detect the time lags of the variables, according to which all variables are aligned. Then, CA is performed to extract the LTER among the non-stationary variables and the monitoring statistics are constructed by LTER and the stationary variables to determine whether the fault occurs in the process. The proposed monitoring strategy is validated by a numerical case and an actual industrial case, and compared with other non-stationary process monitoring methods, the results demonstrate the effectiveness of CA considered time lags.