Isolation of Overtemperature Fault in an Industrial Boiler Using Tree-Structured Sparsity-Based Reconstruction

Abstract

This paper develops a data-driven method for the isolation of overtemperature fault in an industrial boiler using the tree-structured sparsity technique, which relies on reconstruction-based contribution of principal component analysis (PCA). Unlike conventional fault tree analysis-based approaches, the developed method incorporates fault tree structure into the reconstruction objective of PCA in the form of tree-structured sparsity regularization term so that data model and process structure can be integrated. The tree-structured sparsity regularization term allows the selection of fault variables in a group-wise way along the fault tree, resulting in sparse and interpretable fault isolation. For the purpose of online application, this paper further designed an updating procedure for the regularization parameter and introduced a fast alternating multiplier method (FADMM) for parameter optimization. Application results on real data from an industrial boiler show that the proposed method outperforms competitive methods. In addition, by virtue of the fault tree structure, the developed method is able to provide useful information for root cause diagnosis and fault propagation analysis.