Euclidean-Riemannian Joint Low-Rank Projections for Industrial Process Monitoring

Abstract

Due to industrial processes usually characterized by multiple operating units and complex interactions, process data is formulated with hybrid structures, indicating that Euclidean and non-Euclidean structures simultaneously exist among data, which poses a great challenge for process monitoring. However, existing monitoring methods tend to analyze data in terms of a single subject in which a single Euclidean metric or non-Euclidean metric is employed to represent data, which probably deteriorates the model accuracy. This paper proposes a novel method called Euclidean-Riemannian joint low-rank projections for process monitoring. A multiple structure embedding-guided learning framework, in which Euclidean space and Riemannian manifold are mapped into a common subspace, is developed to exploit the underlying information on heterogeneous data spaces. Furthermore, the low-rank constraint is exploited to alleviate the negative influence of corruption so that monitoring results become more reliable. The l 2,1 norm is forced on projection matrix, which enables the proposed approach to be more flexible in the selection of useful information. By this means, the reduced-dimensional representations captured can give more insights into the intrinsic information on data, enhancing the fault detection capability. The experimental results on the Tennessee benchmark platform and the real-world industrial processes, namely fluidized catalytic cracking process, show the effectiveness of our proposed approach.