Industrial process monitoring with density-based kernel structure preserving projections model

Abstract

Practice shows that the influence of latent geometric structures and nonlinear relationships in the industrial data on process monitoring cannot be ignored. Traditional methods focus on the distance information of a pair of points, while we think that the density information of more points can better reflect the geometry shape of the data. Based on this idea, this paper further combines density information with distance information to preserve data latent structure information, a nonlinear process monitoring method to extract global-local spatial geometric structure information is proposed, called density-based kernel structure preserving projections (DKSPP). The principle of this method is: First, search for the neighbors and non-neighbors of a sample in the high-dimensional feature space after nonlinear mapping. Then, the spatial distance information and density information among the data points are comprehensively constructed, and the projection points after dimensionality reduction are constrained to achieve global-local structure preservation. Finally, the obtained projections are used to set anomaly detection indices, and a DKSPP-based anomaly detection model is implemented to monitor processes. The experimental results of a benchmark dataset and an actual industrial application demonstrate the high detection rates and performance of the proposed method.