(345f) Feature Extraction for Process Monitoring with Nonstationary Measurements Based on Principal Component Analysis and Mutual Information

Data-driven process monitoring technology has received considerable attention in chemical industrial process with the rapid development of data analytics. In reality, production loads in real industrial processes are frequently adjusted due to market situation and administration regulation, which makes some process variables nonstationary in nature. These time-varying characteristics in mean, variance or variable correlation violate the idealized assumptions in traditional multivariate statistics such as principal component analysis (PCA), which has been the most commonly applied multivariate projection method. One of the challenges is that process data operated under normal conditions could not be projected to a fixed normalization center by PCA, and therefore the normal fluctuation may be considered as a faulty deviation. The consequence is that whether the process deviation cannot be early identified or the normal fluctuations are considered to be a fault. The other challenge is that the time-varying characteristic in variable correlation cannot be completely captured by PCA, which means the faulty deviation may be buried in the nonstationary trends of process variables and cannot be detected timely.

In this work, a feature extraction for process monitoring with nonstationary measurements based on principal component analysis and mutual information is proposed to address this issue. Process variables are divided into two parts, stationary variables and nonstationary variables, by Augmented Dickey-Fuller (ADF) test. PCA method is applied to detect the fault introduced to stationary process variables group. For another part, variable correlation is calculated by mutual information. Features of the mutual information matrix are extracted by singular value decomposition. As the essential relationship among variables will not be changed by nonstationary characteristics caused by normal fluctuation, angle between is similar. Therefore, the cosine similarity is utilized as a test statistic to measure this similarity. Once a faulty deviation occurs, mutual information matrix is significantly different from that under normal conditions because variable correlation is influenced by the fault, and corresponding cosine similarity statistic will exceed the threshold established from normal conditions, by which faults introduced to variables with time varying characteristics can be also detected. The proposed method is applied to both Tennessee Eastman process and an industrial process. Better performance is obtained than that by traditional PCA mothed.