Joint Structure Preserving Embedding Model and Its Application for Process Monitoring

Currently, dimensionality reduction methods based on manifold learning are widely applied to industrial process monitoring. However, considering that a number of manifold-based monitoring methods extract only the global geometric information of the original data or the local manifold structure of the neighboring data points, in this article, a novel dimensionality reduction method called joint structure preserving embedding (JSPE) is proposed. The proposed method can preserve the underlying sparse neighbor relations by sparse neighborhood preserving embedding. More importantly, since the geodesic distance can more accurately reflect the distance between two points than the Euclidean distance, the geodesic metric is introduced to discover the intrinsic geometric structure between all pairs of non-neighbor points. Based on the captured global and local information, the latent variables can provide a faithful representation of the original data. Furthermore, a reconstruction-based fault diagnosis strategy is developed under the JSPE-based monitoring framework for locating potential fault variables. Finally, two case studies on numerical examples and the Tennessee Eastman benchmark process are provided to illustrate the validity of the JSPE-based monitoring scheme.