Time Neighborhood Preserving Embedding Model and Its Application for Fault Detection

By incorporating the serial correlations of the process data, a new dimensionality reduction method, named time neighborhood preserving embedding (TNPE) is proposed and applied for dynamic process fault detection. TNPE aims at preserving the local neighborhood structure of the data to learn the underlying geometry manifold. To describe the dynamic characteristic while constructing the neighborhood graph, the search for nearest neighbors is performed with respect to the time sequence adjacent points of each data point. Furthermore, the local dynamic variations of the high-dimensional data are captured by reconstructing each data point from its nearest neighbors. On the basis of the locally linear property of the selected neighbors, the estimated local structure information is robust to the transformations of linear projection. Consequently, by considering the serial correlations, TNPE is able to explore the meaningful dynamic information hidden in high-dimensional data. For fault detection, Hotelling's T-2 and squared prediction error (SPE) are constructed upon the TNPE model. Two case studies on a multivariate dynamic process and the Tennessee Eastman process demonstrate the efficacy of the proposed method.