In this paper, a new fault detection and identification scheme that is based on the global–local structure analysis (GLSA) model is proposed. By exploiting the underlying geometrical manifold and simultaneously keeping the global data information, the GLSA model constructs a dual-objective optimization function for dimension reduction of the process dataset. It combines the advantages of both locality preserving projections (LPP) and principal component analysis (PCA), under a unified framework. Meanwhile, GLSA can successfully avoid the singularity problem that may occur in LPP and shares the orthogonal property of the PCA method. In order to balance the two subobjectives (corresponding to global and local structure preservings), a tuning parameter is introduced, and an energy-function-based strategy is proposed to determine the value of the introduced tuning parameter. For the purpose of fault detection, two statistics are constructed, based on the GLSA model. Furthermore, the Bayesian inference algorithm is introduced upon the two monitoring statistics for fault identification. Two case studies are provided to demonstrate the efficiencies of the GLSA model.

No abstract

In this paper a new dimensionality reduction technique named global-local structure analysis (GLSA) is proposed. It constructs a dual-objective optimization function, which exploits the underlying geometrical manifold and keeps the global information for dimensionality reduction simultaneously. This combines the advantages of locality preserving projections (LPP) and principal component analysis (PCA) under a unified framework. Besides, GLSA successfully avoids the singularity problem in LPP and shares the orthogonal property with PCA. A further contribution of this paper is to propose a strategy for determining the parameter η which is used to balance the subobjectives corresponding to global and local structure preservings. For fault detection purpose, two traditional statistics 2 T and SPE are constructed based on the new proposed GLSA method. Case studies on a numerical example and Tennessee Eastman process demonstrate the efficiencies of GLSA in feature extraction and fault detection.

To monitor industrial processes through a probabilistic manner, the probabilistic principal component analysis (PPCA) method has recently been introduced. However, PPCA has its inherent limitation that it cannot determine the effective dimensionality of latent variables. This paper intends to introduce a Bayesian treatment upon the traditional principal component analysis method for process monitoring, which can automatically determine the effective number of retained principal components. Thus, a Bayesian principal component analysis based monitoring approach is developed. A case study of the Tennessee Eastman (TE) benchmark process shows the feasibility and efficiency of the proposed method.

No abstract