A Nonstationary Process Monitoring Based on Mutual Information among Process Variables

Abstract

Abstract In practical chemical industrial processes, the feed valve is automatically adjusted in response to the control systems and the production load will be adjusted with market situation and administrative regulations. Therefore, process data display nonstationary statistics in practical operation condition and cannot satisfy the ideal assumptions of traditional multivariate statistical methods that process is assumed operating around one preset steady state. Under normal operating conditions, fluctuations or adjustments will only affect the mean and standard deviation of process variables, but the correlation among process variables should follow its inherent mechanism model, whose feature can be statistically captured within certain range. In this paper, a nonstationary process monitoring based on mutual information among process variables is proposed. The Euclidean distance (ED) of eigenvalues of the mutual information matrix under normal operation conditions is calculated to obtain a statistic. Once a fault occurs, the changes in correlation among process variables will be reflected in the mutual information matrix and corresponding ED will exceed the threshold, by which process monitoring can be implemented. A numerical simulation example and a practical cracking process are applied as case studies. The results show a better performance on monitoring nonstationary process than traditional principle component analysis method.