Data-Driven Anomaly Monitoring for Nonstationary Industrial Process

With the changes of set points, external environment, and other factors, industrial processes are always operating under multiple conditions. The variation of the operating conditions easily causes the nonstationary characteristics, which is not conductive to the anomaly monitoring. To solve this problem, data-driven anomaly monitoring method, with the aid of kernel slow feature analysis and Gaussian mixture model, is proposed in this paper. Firstly, a data reorganization strategy is used to identify the changing operating conditions according to the statistical characteristics. Then the static and dynamic data slices and segments can be divided. Secondly, kernel slow feature analysis is introduced to extract the features of these nonlinear data segments. Then both slow and fast features can be obtained. Thirdly, Gaussian mixture model is presented to approximate the non-Gaussian distribution of the extracted slow and fast features. Then the evaluating index of these features, based on the Mahalanobis distance, is utilized to realize the anomaly monitoring. Finally, Tennessee-Eastman benchmark data are applied to evaluate the effectiveness of the proposed anomaly monitoring method. The results demonstrate that the proposed method can accurately monitor the anomalies under nonstationary operating process.